Literature
Display
← All papers

Ambient Diffusion Policy: Imitation Learning from Suboptimal Data in Robotics

Adam Wei · Nicholas Pfaff · Thomas Cohn · Arif Kerem Dayı · Constantinos Daskalakis · Giannis Daras · Russ Tedrake

Preprint posted · arXiv (preprint) ↗

Robotics18 min text · CondensedOriginal paper ↗
Contents

Overview

This edition covers the complete 52-page v1 dated 10 June 2026, including the theoretical proofs, four algorithms and all experimental appendices. The proposed training rule changes when auxiliary data is sampled along the diffusion process; inference retains the ordinary Diffusion Policy procedure.

Complete original figure with every panel, label and legend.
Figure 1: Training Algorithm for Ambient Diffusion Policy. High-quality data can be used during training at all diffusion times. Suboptimal data is restricted to training at low or high diffusion times t∈[0,tmax)∪(tmin,T]t\in[0,t_{\max})\cup(t_{\min},T]. Task visualizations. We evaluate Ambient Diffusion Policy on four types of action suboptimality, six different tasks, and scale to the Open X-Embodiment (OXE) [1].

At a glance

Ambient Diffusion Policy uses a small target dataset together with a larger, imperfect auxiliary dataset. Auxiliary samples contribute only at selected high-noise or low-noise steps, where their global plans or local motion primitives respectively resemble the target data.

Experiments span noisy planning trajectories, a simulated proxy for sim-to-real transfer, incorrect-task demonstrations and heterogeneous OXE mixtures. The strongest OXE tower result averages 15.15 correct blocks versus 11.35 for co-training, approximately 33% more. Evidence for broader robot-data structure is empirical; the formal analysis uses a simplified stationary Gaussian model.

1 Introduction

Large robot datasets mix tasks, embodiments, control quality and environments. Filtering discards useful structure, while training on every part of a mixture can teach its undesirable behavior. Noise-dependent sampling instead asks which features each source can contribute.

Building on Ambient Diffusion Omni and Diffusion Policy, the paper connects low-frequency action structure to early global planning and high-frequency structure to late refinement. It evaluates selective use of auxiliary data across six tasks, including two physical-robot tasks trained with Open X-Embodiment.

2 Related Work

Dataset re-weighting changes how often a source contributes but ordinarily uses it at every noise level. Filtering rejects entire samples, and finetuning can mitigate mixture bias after pretraining. Ambient is complementary to weights and finetuning because it restricts the learning signal within each diffusion trajectory.

Earlier corrupted-data methods often assume a known corruption or discard suspect examples. Quality conditioning, auxiliary dynamics objectives and invariant representations offer other approaches; this paper focuses on action distributions and does not solve arbitrary observation corruption. OpenVLA’s curated mixture supplies an important scaling baseline.

Equation 1
ℒco-training​(θ)=α​ℒ𝒟p​(θ)+(1−α)​ℒ𝒟q​(θ),\mathcal{L}_{\text{co-training}}(\theta)=\alpha\,\mathcal{L}_{\mathcal{D}_{p}}(\theta)+(1-\alpha)\,\mathcal{L}_{\mathcal{D}_{q}}(\theta), (1)

3 Background

The setup distinguishes the desired conditional action distribution from an auxiliary distribution that may contain different low-level behavior or task structure.

3.1 Robot Imitation Learning and Diffusion Policy

A diffusion policy predicts a future action chunk from recent observations by learning denoisers over increasing Gaussian noise levels. Squared-error training ideally recovers the conditional mean of the clean action, which supports reverse-process sampling.

The exposition uses variance-exploding noise for simpler notation. The actual implementation uses variance-preserving diffusion; Appendix B derives the corresponding rescaling, scores and denoisers.

Equations 2–3
At=A0+σ⁡(t)​Z,Z∼𝒩⁡(0,I),A_{t}=A_{0}+\sigma(t)\,Z,\quad Z\sim\mathcal{N}(0,I), (2)
ℒ⁡(θ)=𝔼t,(O,A0),At|A0​[‖hθ​(At,O,t)−A0‖2],\mathcal{L}(\theta)=\mathbb{E}_{t,\,(O,A_{0}),\,A_{t}\mid A_{0}}\left[\|h_{\theta}(A_{t},O,t)-A_{0}\|^{2}\right], (3)

3.2 Problem Setting

The user chooses a small high-quality target set and a larger auxiliary set. Here the target contains expert demonstrations for the desired robot, task and environment. The objective is better target-distribution denoising, rather than faithful imitation of the combined mixture; choosing this partition remains a consequential human design decision.

Original boxed problem statement, including its exact conditional denoising objective.
Selectable problem statement
Problem Statement: Our goal is to leverage the vast pool of suboptimal samples in 𝒟q\mathcal{D}_{q} to improve the estimation of the denoisers, 𝔼p[A0∣At,O]\mathbb{E}_{p}[A_{0}\mid A_{t},O], on the user-defined target distribution.

3.3 Ambient Diffusion Omni

Ambient annotates lower and upper noise cutoffs for auxiliary samples. High noise can erase local discrepancies between target and auxiliary actions, allowing shared global structure to be learned. Low-noise use is separately justified when local primitives agree despite different global tasks.

The retained total-variation contraction bound motivates this idea, but distributional proximity is approximate. The source writes Gaussian covariance with sigma rather than sigma squared immediately after Equation 4; the later formal definitions use sigma squared.

Equation 4
dTV​(pt,qt)≤dTV​(p,q)⋅D2​σt,\displaystyle\mathrm{d}_{\mathrm{TV}}(p_{t},q_{t})\leq\mathrm{d}_{\mathrm{TV}}(p,q)\cdot{D\over 2\sigma_{t}}, (4)

4 Distributional Properties of Robot Data

Across the examined datasets, action power concentrates at low temporal frequencies. The paper links this observation to hierarchical denoising and sensitivity concentrated around neighboring actions. These are measured properties of the tested data, not an empirical survey of every possible robot action distribution.

4.1 Spectral Power Law in Robotics

Power spectral density decays approximately as a frequency power law in OXE and six other action datasets. Gaussian noise masks weak, high-frequency details first, leaving global path structure distinguishable at higher noise. Reverse denoising consequently resolves a broad plan before local collision avoidance and smooth motion.

If auxiliary and target data share that broad plan but differ mainly in jitter or contact details, restricting auxiliary use to sufficiently high noise can preserve its coverage without copying its local defects. Figures 2–4 retain the trajectory, spectrum and sensitivity evidence.

Complete original figure with every panel, label and legend.
Figure 2: The spectral power law induces a global-to-local hierarchy in action diffusion. This is analogous to the “coarse-to-fine” hierarchy in image diffusion [60]. (a) t≥t2t\geq t_{2} (global planning): the policy commits to a high-level path through the maze. t<t2t<t_{2} (local refinement): it refines the plan to be collision-free at t1t_{1} and smooth at t0t_{0}. (b) t3t_{3}: the policy plans to move a block to the right bin. t2t_{2}: it plans which block to grasp. t<t2t<t_{2}: it refines the grasping motion.
Complete original figure with every panel, label and legend.
(a) Spectral power law in robot action data. The power spectral density for OXE [1], a collection of 2.4M robot episodes across 70+ datasets, follows a power law. Figure 11 shows that this spectral power law is a more general property of robot data.
Complete original figure with every panel, label and legend.
(b) Locality in robot action data. Visualizing the sensitivity of the 8th action estimate to its neighboring actions (Eq. (83)). At low noise, the 8th action estimate is most sensitive to its temporal neighbors.
Complete original figure with every panel, label and legend.
Figure 4: The spectral power law makes Gaussian noise act as a high-frequency mask. (a) Noisy trajectory visualizations. (b) PSD analysis for the noisy trajectories in the xx-dimension at each diffusion time: the gray regions indicate masked frequencies, which have low SNR. As tt increases, adding Gaussian noise destroys the signal in the trajectory, starting from the local (high-frequency) features first. This effect is visible in both the noisy trajectories and the PSD plots.

4.2 Locality in Robotics

At low noise, the eighth predicted action is empirically most sensitive to nearby noisy actions. This locality suggests that demonstrations with the wrong task-level ordering may still teach useful grasping or motion primitives when admitted only late in denoising.

The argument requires local agreement. It does not imply that every wrong-task sample is harmless at a chosen low-noise cutoff, or that a learned finite-data denoiser exactly equals the theoretical optimal one.

5 Theoretical Foundations

The exact definition and theorem statements below retain their assumptions. The analysis uses zero-mean stationary Gaussians, whose circulant covariance is diagonalized by the Fourier transform; extension to general robot-data distributions is future work.

Equation
Definition 1 (Spectral Power Law).

Let XX be a zero-mean, stationary random vector on ℝN\mathbb{R}^{N}, and let SX​(f):=𝔼⁡[|(ℱ​X)​(f)|2]S_{X}(f):=\mathbb{E}\big[|(\mathcal{F}X)(f)|^{2}\big] denote its power spectral density in the orthonormal Fourier basis. We say that the distribution of XX follows a spectral power law if there exist a constant C>0C>0, a rolloff exponent α>1\alpha>1, and a cutoff frequency f0≥1f_{0}\geq 1 such that the high-frequency spectrum satisfies:

SX​(f)=C​|f|−αfor all ​|f|≥f0.S_{X}(f)=C|f|^{-\alpha}\quad\text{for all }|f|\geq f_{0}.

The exact spectral behavior below the cutoff f0f_{0} is permitted to be arbitrary but bounded.

5.1 Contraction Through Noise

When low-frequency spectra agree and the remaining spectral tails obey a power-law bound, sufficiently large Gaussian noise bounds KL divergence as the inverse fourth power of noise scale and total variation as the inverse square. The bounds depend on tail amplitude and cutoff, rather than the original total-variation distance.

This theorem assumes exact low-frequency agreement and a specified minimum noise variance. It supports the mechanism under those conditions, rather than guaranteeing useful transfer from an arbitrary auxiliary dataset.

Equation ⋆
Theorem 1 (Spectral Power Law implies fast contraction through noise).

Let p0=𝒩⁡(0,Σp)p_{0}=\mathcal{N}(0,\Sigma_{p}) and q0=𝒩⁡(0,Σq)q_{0}=\mathcal{N}(0,\Sigma_{q}) be zero-mean stationary Gaussians on ℝN\mathbb{R}^{N} with power spectra Sp,SqS_{p},S_{q}. Fix a cutoff frequency f⋆≥1f^{\star}\geq 1, an exponent α>12\alpha>\tfrac{1}{2}, and a constant C>0C>0. Assume:

  1. 1.

    (Low-frequency agreement.) Sp​(f)=Sq​(f)S_{p}(f)=S_{q}(f) for every f≤f⋆f\leq f^{\star}.

  2. 2.

    (Power-law tail.) max⁡(Sp​(f),Sq​(f))≤C​f−α\max(S_{p}(f),S_{q}(f))\leq Cf^{-\alpha} for every ff such that f⋆<f≤⌊N/2⌋f^{\star}<f\leq\lfloor N/2\rfloor.

Let pt:=p0∗𝒩⁡(0,σt2​I)p_{t}:=p_{0}*\mathcal{N}(0,\sigma_{t}^{2}I) and qt:=q0∗𝒩⁡(0,σt2​I)q_{t}:=q_{0}*\mathcal{N}(0,\sigma_{t}^{2}I). If

Original mathematical display with its complete source layout, enclosures, spacing and notation.
Exact source MathML
σt2≥C​(f⋆)−α,\sigma_{t}^{2}\geq C(f^{\star})^{-\alpha}, (⋆\star)

then

Original mathematical display with its complete source layout, enclosures, spacing and notation.
Exact source MathML
dKL(pt∥qt)≤4​C2(2​α−1)​σt4​(f⋆)2​α−1dTV​(pt,qt)≤2​Cσt2​(2​α−1)​(f⋆)2​α−1.\boxed{\mathrm{d}_{\mathrm{KL}}(p_{t}\|q_{t})\leq\frac{4C^{2}}{(2\alpha-1)\sigma_{t}^{4}(f^{\star})^{2\alpha-1}}}\qquad\boxed{\mathrm{d}_{\mathrm{TV}}(p_{t},q_{t})\leq\frac{\sqrt{2}C}{\sigma_{t}^{2}\sqrt{(2\alpha-1)(f^{\star})^{2\alpha-1}}}}.

5.2 Spectral Power Law Implies Locality

The stated locality result bounds the change in one denoised coordinate after masking distant input coordinates by a term proportional to signal length and amplitude and inversely proportional to neighborhood radius. The appendix derives spatial decay through the Wiener filter.

The displayed final bound has no explicit noise dependence. Moreover, the proof uses a smooth power law from frequency one, whereas Definition 1 permits an arbitrary bounded low-frequency region. These source assumptions and the unchanged bound should be considered when interpreting the broader low-noise locality claim.

Equation
Theorem 2 (Spectral Power Law implies locality of the optimal denoiser).

Let X0∼𝒩⁡(0,Σ)X_{0}\sim\mathcal{N}(0,\Sigma) be a zero-mean stationary Gaussian on ℝN\mathbb{R}^{N} whose power spectrum S⁡(f)S(f) follows a spectral power law with constant C>0C>0 and exponent α>1\alpha>1. Let Xt=X0+σt​ZX_{t}=X_{0}+\sigma_{t}Z be the variance-exploding noisy observation, and ht⋆​(x):=𝔼⁡[X0∣Xt=x]h^{\star}_{t}(x):=\mathbb{E}[X_{0}\mid X_{t}=x] denote the MMSE optimal denoiser.

Let MiM_{i} be a circulant mask that sets to zero all coordinates of the input xx at a circular distance strictly greater than LL from index ii. Then, for any mask size 1≤L≤⌊N/2⌋1\leq L\leq\lfloor N/2\rfloor, the absolute error in the ii-th coordinate between the full optimal denoiser and the masked denoiser is bounded by:

Original mathematical display with its complete source layout, enclosures, spacing and notation.
Exact source MathML
|(ht⋆​(x))i−(ht⋆​(Mi​x))i|≤α​N​‖x‖∞8​L.\boxed{\bigg|(h^{\star}_{t}(x))_{i}-(h^{\star}_{t}(M_{i}x))_{i}\bigg|\leq\frac{\alpha N\|x\|_{\infty}}{8L}.}

6 Ambient Diffusion Policy: Method

Training has two phases: determine usable noise intervals for auxiliary data, then sample only admissible data at each randomly drawn diffusion time.

6.1 Phase 1: Dataset annotation

A sweep can choose the cutoffs. Alternatively, train a balanced classifier to distinguish noisy target actions from noisy auxiliary actions and choose the first noise level where its mean prediction on auxiliary data exceeds a threshold near one half.

Labels may be assigned per dataset, per trajectory, per partition or per action chunk. The theoretical guarantee assumes a near-optimal population classifier; finite-data training does not establish that assumption. The classifier here operates on actions and time, so marginal action alignment does not by itself certify observation-conditioned alignment.

Equation 5
tmin=inf{t:𝔼(O,A0)∼q,At|A0​[cϕ∗​(At,t)]>0.5−τ}.t_{\min}=\inf\{t:\mathbb{E}_{(O,A_{0})\sim q,\,A_{t}|A_{0}}[c_{\phi^{*}}(A_{t},t)]>0.5-\tau\}. (5)

6.2 Phase 2: Training

First sample diffusion times uniformly; then sample target data or auxiliary data whose allowed intervals include each time. Target demonstrations remain admissible throughout. The standard denoising objective and inference procedure can remain unchanged.

An optional Ambient loss predicts through an intermediate noisy sample, and a classifier rejection sampler is proposed separately. Most experiments use the ordinary loss; the rejection-sampling variant is theoretically motivated but experimentally untested.

7 Controlled Experiments

Controlled experiments isolate noisy trajectories, simulator mismatch and incorrect task logic. Baselines include target-only filtering, weighted co-training and finetuning; the appendix retains data generation, selection and evaluation details.

7.1 Noisy Trajectories

A two-dimensional maze uses 50 smooth GCS trajectories and 5,000 jittery RRT trajectories. Across 1,000 trials, Ambient and co-training reach similar success, 99.5% and 99.4%, while Ambient’s squared-acceleration metric is 31.0 versus 62.2. Target-only training succeeds 57.5% of the time.

Using only auxiliary data above the cutoff and only target data below it still reaches 99.0% success with smoothness 29.9. This supports the separation of global coverage and local motion quality in this task. Smoothness statistics include successful rollouts only.

Complete original figure with every panel, label and legend.
Figure 5: GCS (left) vs RRT (right) data. The high-quality GCS [65] trajectories are smoother and shorter than their RRT [66] counterparts.
Table 1: Maze and Neural Motion Planning results (1000 trials each). In both settings, Ambient Diffusion Policy achieves the highest success rate and is substantially smoother than co-training.
Training Algo. Success Rate ↑\uparrow
Smoothness ↓\downarrow
(Eq. (84))
Data Filtering 57.5−3.1+3.157.5^{+3.1}_{-3.1} % 31.931.9
Co-train
(α∗=0.019\alpha^{*}=0.019)
99.4−0.7+0.499.4^{+0.4}_{-0.7} % 62.262.2
Ambient
(σtmin∗=0.074\sigma_{t_{\min}^{*}}=0.074)
99.5−0.7+0.3\mathbf{99.5^{+0.3}_{-0.7}} % 31.0\mathbf{31.0}
(a) 2D maze.
Training Algo. Success Rate ↑\uparrow
Smoothness ↓\downarrow
(Eq. (84))
Data Filtering 46.0−3.1+3.146.0^{+3.1}_{-3.1} % 3.9\mathbf{3.9}
Co-train
(α∗=0.091\alpha^{*}=0.091)
59.9−3.1+3.159.9^{+3.1}_{-3.1} % 42.742.7
Ambient
(σtmin∗=0.025\sigma_{t_{\min}^{*}}=0.025)
65.9−3.0+2.9\mathbf{65.9^{+2.9}_{-3.0}} % 31.431.4
(b) 7-DoF neural motion planning.
Complete original figure with every panel, label and legend.
Figure 6: The “smoothness vs success rate” Pareto frontier for Ambient (blue) dominates co-training (red) in both the (a) maze and (b) neural motion planning experiments. This trade-off is controlled by σtmin\sigma_{t_{\min}} for Ambient and α\alpha for co-training.
Selectable original panel labels

(a) 2D maze.

(b) 7-DoF neural motion planning.

7.2 Neural Motion Planning

The seven-joint planning task uses 100,000 optimized trajectories and one million BiRRT trajectories. For one predicted full plan per query, Ambient achieves 65.9% success versus 59.9% for co-training, with squared acceleration 31.4 versus 42.7; filtering is smoother but succeeds only 46.0%.

These are open-loop simulated plans. Evaluation accepts up to 2 cm penetration and endpoint error within a .2-radian ball; its success metric therefore does not establish strict collision-free execution on hardware. Best-of-100 selection is reported separately.

7.3 Sim-to-Real Distribution Shift

Planar pushing uses 50 target-simulator teleoperation demonstrations and 2,000 source-simulator planned demonstrations from the earlier sim-and-real study. The two simulators differ visually, dynamically and in action generation, providing a controlled proxy for a sim-to-real gap.

Across 200 trials, per-datapoint cutoff annotation reaches 93.5% success versus 84.5% for co-training and 56.5% for filtering. Per-dataset annotation reaches 87.0%; adding low-noise auxiliary use gives 92.0%, consistent with little benefit when local contact dynamics differ.

Table 2: Main planar pushing results (200 trials). All Ambient variations outperform co-training [12].
Training Algo. Data Filtering Co-training [12]
Ambient
(tmint_{\min} per dataset)
Ambient
(tmint_{\min} per datapoint)
Ambient + Locality
(σtmax=0.025\sigma_{t_{\max}}=0.025)
Success Rate ↑\uparrow 56.5−7.2+7.056.5^{+7.0}_{-7.2} % 84.5−5.8+4.784.5^{+4.7}_{-5.8} % 87.0−5.5+4.387.0^{+4.3}_{-5.5} % 93.5−4.4+3.0\mathbf{93.5^{+3.0}_{-4.4}} % 92.0−4.7+3.492.0^{+3.4}_{-4.7} %

7.4 Task Mismatch

The target set contains 50 demonstrations sorting colors into the correct bins, while 200 auxiliary demonstrations reverse the rule. Low-noise-only auxiliary use reaches 93.3% correctly placed blocks versus 22.7% for co-training and 61.0% for filtering, over about 800 blocks.

Motion success counts blocks placed in either bin; logic accuracy is conditional on a successful placement. Their product is overall correctness. Ambient retains 98.2% conditional logic and 95.0% placement success; task-conditioning and disjoint-noise ablations appear in Appendix G.4.

The text says to set the lower cutoff to zero while isolating low-noise use, which conflicts with its union-of-intervals rule: a zero lower cutoff would also admit all positive high-noise times. The described experiment instead uses auxiliary data only below the upper cutoff; the source wording is retained as a specification inconsistency.

Table 3: Main block sorting results (∼\sim800 blocks). Only Ambient Diffusion Policy (Locality) performs well on all metrics.
Training Algo.
Data Filtering
𝒟q\mathcal{D}_{q} Only
Co-train
(α∗=0.9\alpha^{*}=0.9)
Locality
(σtmax∗=0.46\sigma_{t_{\max}^{*}}=0.46)
Logic ↑\uparrow 98.6−1.5+0.8\mathbf{98.6^{+0.8}_{-1.5}} % 3.0−1.2+1.63.0^{+1.6}_{-1.2} % 26.0−3.2+3.426.0^{+3.4}_{-3.2} % 98.2−1.2+0.898.2^{+0.8}_{-1.2} %
Motion ↑\uparrow 61.9−3.5+3.461.9^{+3.4}_{-3.5} % 83.0−2.8+2.583.0^{+2.5}_{-2.8} % 87.2−2.5+2.287.2^{+2.2}_{-2.5} % 95.0−1.7+1.4\mathbf{95.0^{+1.4}_{-1.7}} %
Success Rate ↑\uparrow 61.0−3.5+3.461.0^{+3.4}_{-3.5} % 2.5−1.0+1.32.5^{+1.3}_{-1.0} % 22.7−2.9+3.122.7^{+3.1}_{-2.9} % 93.3−2.0+1.6\mathbf{93.3^{+1.6}_{-2.0}} %
Complete original figure with every panel, label and legend.
Figure 7: Left: Performance metrics vs σtmax.\sigma_{t_{\max}}. The logic metric plateaus before σtmax∗=0.46\sigma_{t_{\max}^{*}}=0.46 and deteriorates rapidly afterwards; the opposite is true for the motion metric. Thus, the policy learns local motion-level features for t∈[0,tmax∗)t\in[0,t_{\max}^{*}) and global logic-level features for t∈(tmax∗,T]t\in(t_{\max}^{*},T]. Right: Unlike co-training, Ambient achieves a near-optimal trade-off between logic and motion metrics.

7.5 Finetuning Comparison

Finetuning the best co-trained and Ambient checkpoints on target data generally favors starting from Ambient. The unfinetuned Ambient policy can already outperform a finetuned co-trained policy, and further target-only finetuning does not consistently help Ambient. These comparisons include checkpoint selection, as detailed in Appendix I.

Complete original figure with every panel, label and legend.
Figure 8: Left: success rate of policies finetuned from different base policies. Right: smoothness of finetuned policies in the maze experiment. We compare smoothness instead of success rate since all policies achieved 99.0%+ success rate.

8 Scaling Experiments: Open X-Embodiment

The physical-robot scaling tasks are drawer-based table cleaning and building alternating-color, alternating-orientation block towers. Target sets contain 50 or 150 cleaning demonstrations, or 35 tower demonstrations. Auxiliary data expands from 27-dataset Magic Soup++ to a 48-dataset custom OXE subset.

Per-source weights are renormalized among datasets admissible at each noise time. Co-training keeps its mixture fixed over time; Ambient changes the mixture as source cutoffs become active. This makes source weighting and interval selection distinct controls.

Equation 6
w~t(i)={wi∑j∈Stwjif ​i∈St0otherwise.\tilde{w}_{t}^{(i)}=\begin{cases}\dfrac{w_{i}}{\sum_{j\in S_{t}}w_{j}}&\text{if }i\in S_{t}\\ \quad\quad 0&\text{otherwise}\end{cases}. (6)

8.1 Main Results

With 50 cleaning demonstrations, Figure 9 shows approximately .83 completion for the strongest custom-OXE Ambient variant versus .71 for co-training and .68 for filtering. With 150 demonstrations, the strongest plotted Ambient score is .87 versus .82 for custom-OXE co-training; the value of auxiliary data is smaller.

Tower building averages 15.15 correct blocks for Ambient with locality, versus 11.35 for co-training and 8.20 for filtering, about 33% and 85% relative increases. Each policy receives 20 trials. Percentage changes in the prose mix completion-point differences and relative tower-height changes; the original plots retain the actual values.

Complete original figure with every panel, label and legend.
Figure 9: When scaled to OXE, Ambient Diffusion Policy outperforms data filtering and co-training by up to 15% on table cleaning and 84% on tower building (20 trials per policy).

8.2 Ablations

Removing dataset weights lowers the tested Ambient cleaning scores by roughly .03–.09. Unweighted co-trained policies were judged too unsafe to evaluate on hardware, so their plot entries are unavailable, not measured zero success.

Additional measurements cover grasp attempts, time per object and finetuning. Reduced sensitivity in these experiments is useful practical evidence, but does not imply arbitrary mixture weights are safe or irrelevant.

Complete original figure with every panel, label and legend.
Figure 10: Ambient is significantly less sensitive to dataset re-weighting than co-training. The Ambient Diffusion Policy performed at most 9% worse without re-weighting. The co-trained policies were too dangerous to evaluate without re-weighting.

9 Limitations and Future Work

Observation-space shifts remain unresolved. Cutoff classifiers are imperfect and can lose to expensive sweeps; the Gaussian theory is limited, and training on more data costs more than filtering. Choosing the target/auxiliary partition remains essential.

Generalist pretraining would require a principled target definition beyond one task and environment. Better annotation, broader theory and learning from imperfect finetuning data are proposed directions.

10 Conclusion

The central contribution is noise-dependent access to auxiliary demonstrations. The tested cases show that retaining useful global or local structure can improve the trade-off between task success and behavior quality, while the appendix records the assumptions, implementation choices and negative results that limit that conclusion.

Acknowledgments

The authors acknowledge discussions and hardware assistance, NSF and NSERC graduate support, and MIT Supercloud compute.

Appendix A Ambient Diffusion: Theoretical Foundations

This appendix proves the two Gaussian spectral results and analyzes the classifier threshold. All displayed derivations and exact theorem/proposition statements remain available below.

Preliminaries.

The preliminaries define total variation, chi-squared divergence and a LeCam quantity. There is a normalization inconsistency: the printed total-variation definition lacks the one-half factor used later in the rejection-sampling proof, while the stated Pinsker constant follows the latter convention. The original definitions and inequalities are preserved.

Equations
dTV​(P,Q)\displaystyle\mathrm{d}_{\mathrm{TV}}(P,Q) =𝔼Q​|1−d​Pd​Q|,\displaystyle=\mathbb{E}_{Q}\!\left|1-\frac{dP}{dQ}\right|,
χ2(P∥Q)\displaystyle\chi^{2}(P\|Q) =𝔼Q​(d​Pd​Q)2−1,\displaystyle=\mathbb{E}_{Q}\!\left(\frac{dP}{dQ}\right)^{2}-1,
LC⁡(P,Q)\displaystyle\mathrm{LC}(P,Q) =2χ2(P∥12P+12Q).\displaystyle=2\chi^{2}\!\left(P\,\big\|\,\tfrac{1}{2}P+\tfrac{1}{2}Q\right).
dTV​(P,Q)2≤12​KL​(P,Q)\mathrm{d}_{\mathrm{TV}}(P,Q)^{2}\leq\tfrac{1}{2}\mathrm{KL}(P,Q)
dTV​(P,Q)2≤12​LC​(P,Q)\mathrm{d}_{\mathrm{TV}}(P,Q)^{2}\leq\tfrac{1}{2}\mathrm{LC}(P,Q)

A.1 Spectral Power Law implies fast contraction through noise

Fourier diagonalization reduces Gaussian KL divergence to a sum of spectral variance ratios. Low-frequency agreement removes those terms; the noise lower bound keeps each remaining ratio at least one half, allowing a quadratic bound on the logarithmic expression.

Summing the squared power-law tail and bounding it by an integral yields the inverse-fourth-power noise dependence. Pinsker then gives the stated total-variation estimate. The derivation relies on stationarity, spectral agreement and the tail/noise assumptions throughout.

Equations 7–35
dKL(pt∥qt)=dKL(ℱ#pt∥ℱ#qt).\displaystyle\mathrm{d}_{\mathrm{KL}}(p_{t}\|q_{t})=\mathrm{d}_{\mathrm{KL}}(\mathcal{F}\#p_{t}\|\mathcal{F}\#q_{t}). (7)
Σℱ​#​pt=𝔼x∼pt​[(F​x)​(F​x)∗]\displaystyle\Sigma_{\mathcal{F}\#p_{t}}=\mathbb{E}_{x\sim p_{t}}\left[(Fx)(Fx)^{*}\right] (8)
=F​𝔼x∼pt​[x​xT]​F∗\displaystyle=F\mathbb{E}_{x\sim p_{t}}[xx^{T}]F^{*} (9)
=F​Σpt​F∗.\displaystyle=F\Sigma_{p_{t}}F^{*}. (10)
Σpt=F∗​(Λp+σt2​I)​F.\displaystyle\Sigma_{p_{t}}=F^{*}(\Lambda_{p}+\sigma^{2}_{t}I)F. (11)
Σℱ​#​pt=F​F∗​(Λp+σt2​I)​F​F∗\displaystyle\Sigma_{\mathcal{F}\#p_{t}}=FF^{*}(\Lambda_{p}+\sigma^{2}_{t}I)FF^{*} (12)
=Λp+σt2​I.\displaystyle=\Lambda_{p}+\sigma^{2}_{t}I. (13)
dKL(pt∥qt)≤12∑i=0N−1(Λp)i,i+σt2(Λq)i,i+σt2−ln((Λp)i,i+σt2(Λq)i,i+σt2)−1\displaystyle\mathrm{d}_{\mathrm{KL}}(p_{t}\|q_{t})\leq\frac{1}{2}\sum_{i=0}^{N-1}\frac{(\Lambda_{p})_{i,i}+\sigma_{t}^{2}}{(\Lambda_{q})_{i,i}+\sigma_{t}^{2}}-\mathrm{ln}\left(\frac{(\Lambda_{p})_{i,i}+\sigma_{t}^{2}}{(\Lambda_{q})_{i,i}+\sigma_{t}^{2}}\right)-1 (14)
=12​∑f=0N−1Sp​(f)+σt2Sq​(f)+σt2−ln⁡(Sp​(f)+σt2Sq​(f)+σt2)−1\displaystyle=\frac{1}{2}\sum_{f=0}^{N-1}\frac{S_{p}(f)+\sigma_{t}^{2}}{S_{q}(f)+\sigma_{t}^{2}}-\mathrm{ln}\left(\frac{S_{p}(f)+\sigma_{t}^{2}}{S_{q}(f)+\sigma_{t}^{2}}\right)-1 (15)
≤∑f>f⋆⌊N/2⌋Sp​(f)+σt2Sq​(f)+σt2−ln⁡(Sp​(f)+σt2Sq​(f)+σt2)−1,\displaystyle\leq\sum_{f>f^{\star}}^{\lfloor N/2\rfloor}\frac{S_{p}(f)+\sigma_{t}^{2}}{S_{q}(f)+\sigma_{t}^{2}}-\mathrm{ln}\left(\frac{S_{p}(f)+\sigma_{t}^{2}}{S_{q}(f)+\sigma_{t}^{2}}\right)-1, (16)
h⁡(r)=g⁡(r)−(r−1)2\displaystyle h(r)=g(r)-(r-1)^{2} (17)
=r−ln⁡r−1−r2+2​r−1\displaystyle=r-\ln r-1-r^{2}+2r-1 (18)
=−r2+3​r−ln⁡r−2.\displaystyle=-r^{2}+3r-\ln r-2. (19)
h′​(r)=−2​r+3−1r\displaystyle h^{\prime}(r)=-2r+3-\frac{1}{r} (20)
=−2​r2+3​r−1r.\displaystyle=\frac{-2r^{2}+3r-1}{r}. (21)
r1=12,r2=1.\displaystyle r_{1}=\frac{1}{2},\quad r_{2}=1. (22)
Sp​(f)+σt2Sq​(f)+σt2≥σt2max⁡(Sp​(f),Sq​(f))+σt2.\displaystyle\frac{S_{p}(f)+\sigma_{t}^{2}}{S_{q}(f)+\sigma_{t}^{2}}\geq\frac{\sigma_{t}^{2}}{\max(S_{p}(f),S_{q}(f))+\sigma_{t}^{2}}. (23)
max⁡(Sp​(f),Sq​(f))≤C​f−α,\displaystyle\max(S_{p}(f),S_{q}(f))\leq Cf^{-\alpha}, (24)
Sp​(f)+σt2Sq​(f)+σt2≥σt2C​f−α+σt2.\displaystyle\frac{S_{p}(f)+\sigma_{t}^{2}}{S_{q}(f)+\sigma_{t}^{2}}\geq\frac{\sigma_{t}^{2}}{Cf^{-\alpha}+\sigma_{t}^{2}}. (25)
σt2≥C​1(f⋆)α.\displaystyle\sigma_{t}^{2}\geq C\frac{1}{(f^{\star})^{\alpha}}. (26)
Cfα<C(f⋆)α,∀f​ such that ​f⋆<f≤⌊N/2⌋.\displaystyle\frac{C}{f^{\alpha}}<\frac{C}{(f^{\star})^{\alpha}},\quad\forall f\text{ such that }f^{\star}<f\leq\lfloor N/2\rfloor. (27)
C​f−α<σt2.\displaystyle Cf^{-\alpha}<\sigma_{t}^{2}. (28)
Sp​(f)+σt2Sq​(f)+σt2≥1/2.\displaystyle\frac{S_{p}(f)+\sigma_{t}^{2}}{S_{q}(f)+\sigma_{t}^{2}}\geq 1/2. (29)
dKL(pt∥qt)≤∑f>f⋆⌊N/2⌋(Sp​(f)−Sq​(f)Sq​(f)+σt2)2\displaystyle\mathrm{d}_{\mathrm{KL}}(p_{t}\|q_{t})\leq\sum_{f>f^{\star}}^{\lfloor N/2\rfloor}\left(\frac{S_{p}(f)-S_{q}(f)}{S_{q}(f)+\sigma_{t}^{2}}\right)^{2} (30)
≤∑f>f⋆⌊N/2⌋(2​max⁡(Sp​(f),Sq​(f))σt2)2\displaystyle\leq\sum_{f>f^{\star}}^{\lfloor N/2\rfloor}\left(\frac{2\max(S_{p}(f),S_{q}(f))}{\sigma_{t}^{2}}\right)^{2} (31)
≤∑f>f⋆⌊N/2⌋4​C2​f−2​ασt4\displaystyle\leq\sum_{f>f^{\star}}^{\lfloor N/2\rfloor}4\frac{C^{2}f^{-2\alpha}}{\sigma_{t}^{4}} (32)
=4​C2σt4​∑f>f⋆⌊N/2⌋(1f)2​α\displaystyle=\frac{4C^{2}}{\sigma_{t}^{4}}\sum_{f>f^{\star}}^{\lfloor N/2\rfloor}\left(\frac{1}{f}\right)^{2\alpha} (33)
≤4​C2σt4​∫f⋆∞(1f)2​α​𝑑f\displaystyle\leq\frac{4C^{2}}{\sigma_{t}^{4}}\int_{f^{\star}}^{\infty}\left(\frac{1}{f}\right)^{2\alpha}\mathrm{d}f (34)
=4​C2σt4​(2​α−1)​(f⋆)2​α−1.\displaystyle=\frac{4C^{2}}{\sigma_{t}^{4}(2\alpha-1)(f^{\star})^{2\alpha-1}}. (35)

A.2 Spectral Power Law implies locality

The Gaussian conditional mean is a circulant Wiener filter. Masking distant coordinates leaves a convolution tail; two finite differences in frequency relate that tail to the total variation of the filter’s slope. A power-law slope estimate gives inverse-square kernel decay and an inverse-radius masking-error bound.

The proof treats a smooth power law from frequency one, without separately bounding the arbitrary low-frequency part permitted by the definition. Its final upper bound also retains a factor of signal length and omits explicit noise dependence. The edition preserves the argument as published rather than strengthening its scope.

Equations 36–58
𝔼⁡[X0|Xt=x]=Σ​(Σ+σt2​I)−1​x.\displaystyle\mathbb{E}[X_{0}|X_{t}=x]=\Sigma(\Sigma+\sigma_{t}^{2}I)^{-1}x. (36)
(𝔼⁡[X0|Xt=x])i=∑j=0N−1kt​((i−j)modN)​xj,\displaystyle(\mathbb{E}[X_{0}|X_{t}=x])_{i}=\sum_{j=0}^{N-1}k_{t}((i-j)\mod N)x_{j}, (37)
errori=|∑j=0N−1kt​((i−j)modN)​xj\displaystyle\mathrm{error}_{i}=\bigg|\sum_{j=0}^{N-1}k_{t}((i-j)\mod N)x_{j}
−∑j=0N−1kt((i−j)modN)Mixj|\displaystyle-\sum_{j=0}^{N-1}k_{t}((i-j)\mod N)M_{i}x_{j}\bigg| (38)
=|∑j:dist⁡(i,j)>Lkt((i−j)modN)xj|,\displaystyle=\bigg|\sum_{j:\mathrm{dist}(i,j)>L}k_{t}((i-j)\mod N)x_{j}\bigg|, (39)
≤||x||∞∑j:dist⁡(i,j)>L|kt((i−j)modN)|,\displaystyle\leq||x||_{\infty}\sum_{j:\mathrm{dist}(i,j)>L}\bigg|k_{t}((i-j)\mod N)\bigg|, (40)
errori≤‖x‖∞​∑m=L+1⌊N/2⌋|kt​(m)|+|kt​(N−m)|.\displaystyle\mathrm{error}_{i}\leq||x||_{\infty}\sum_{m=L+1}^{\lfloor N/2\rfloor}|k_{t}(m)|+|k_{t}(N-m)|. (41)
errori≤2​‖x‖∞​∑m=L+1⌊N/2⌋|kt​(m)|.\displaystyle\mathrm{error}_{i}\leq 2||x||_{\infty}\sum_{m=L+1}^{\lfloor N/2\rfloor}|k_{t}(m)|. (42)
kt​(m)=(Σ​(Σ+σt2​I)−1​e1)​(m).\displaystyle k_{t}(m)=(\Sigma(\Sigma+\sigma_{t}^{2}I)^{-1}e_{1})(m). (43)
kt​(m)=(F∗​W​F​e1)​(m),\displaystyle k_{t}(m)=(F^{*}WFe_{1})(m), (44)
kt​(m)=1N​∑f=0N−1Wf,f​ei​2​π​f​mN.\displaystyle k_{t}(m)=\frac{1}{N}\sum_{f=0}^{N-1}W_{f,f}e^{i\frac{2\pi fm}{N}}. (45)
kt​(m)=1N​∑f=0N−1Wf​cos⁡(2​π​f​mN).\displaystyle k_{t}(m)=\frac{1}{N}\sum_{f=0}^{N-1}W_{f}\cos\left(\frac{2\pi fm}{N}\right). (46)
2​(1−cos⁡ωm)​kt​(m)\displaystyle 2(1-\cos\omega_{m})k_{t}(m)
=1N​∑f=0N−1Wf​[2​cos⁡(f​ωm)−2​cos⁡(f​ωm)​cos⁡(ωm)].\displaystyle=\frac{1}{N}\sum_{f=0}^{N-1}W_{f}\left[2\cos(f\omega_{m})-2\cos(f\omega_{m})\cos(\omega_{m})\right]. (47)
2​(1−cos⁡ωm)​kt​(m)\displaystyle 2(1-\cos\omega_{m})k_{t}(m)
=1N∑f=0N−1Wf[2cos(fωm)−cos((f+1)ωm)\displaystyle=\frac{1}{N}\sum_{f=0}^{N-1}W_{f}\Big[2\cos(f\omega_{m})-\cos((f+1)\omega_{m})
−cos((f−1)ωm)].\displaystyle-\cos((f-1)\omega_{m})\Big]. (48)
2(1−cosωm)kt(m)=−1N∑f=0N−1(Δ2Wf)cos(fωm).\displaystyle 2(1-\cos\omega_{m})k_{t}(m)=-\frac{1}{N}\sum_{f=0}^{N-1}(\Delta^{2}W_{f})\cos(f\omega_{m}). (49)
4​sin2⁡(π​mN)​|kt​(m)|≤1N​∑f=0N−1|Δ2​Wf|.\displaystyle 4\sin^{2}\left(\frac{\pi m}{N}\right)|k_{t}(m)|\leq\frac{1}{N}\sum_{f=0}^{N-1}|\Delta^{2}W_{f}|. (50)
W′​(f)=−α​C​σt2​fα−1(C+σt2​fα)2.\displaystyle W^{\prime}(f)=\frac{-\alpha C\sigma_{t}^{2}f^{\alpha-1}}{(C+\sigma_{t}^{2}f^{\alpha})^{2}}. (51)
|W′​(f)|≤α​C​σt2​fα−14​C​σt2​fα=α4​f≤α4.\displaystyle|W^{\prime}(f)|\leq\frac{\alpha C\sigma_{t}^{2}f^{\alpha-1}}{4C\sigma_{t}^{2}f^{\alpha}}=\frac{\alpha}{4f}\leq\frac{\alpha}{4}. (52)
∑f=0N−1|Δ2​Wf|≤α.\displaystyle\sum_{f=0}^{N-1}|\Delta^{2}W_{f}|\leq\alpha. (53)
4​sin2⁡(π​mN)​|kt​(m)|≤αN.\displaystyle 4\sin^{2}\left(\frac{\pi m}{N}\right)|k_{t}(m)|\leq\frac{\alpha}{N}. (54)
4​(4​m2N2)​|kt​(m)|≤αN⟹|kt​(m)|≤α​N16​m2.\displaystyle 4\left(\frac{4m^{2}}{N^{2}}\right)|k_{t}(m)|\leq\frac{\alpha}{N}\implies|k_{t}(m)|\leq\frac{\alpha N}{16m^{2}}. (55)
errori≤2​‖x‖∞​∑m=L+1⌊N/2⌋|kt​(m)|\displaystyle\mathrm{error}_{i}\leq 2\|x\|_{\infty}\sum_{m=L+1}^{\lfloor N/2\rfloor}|k_{t}(m)|
≤α​N​‖x‖∞8​∑m=L+1⌊N/2⌋1m2.\displaystyle\leq\frac{\alpha N\|x\|_{\infty}}{8}\sum_{m=L+1}^{\lfloor N/2\rfloor}\frac{1}{m^{2}}. (56)
∑m=L+1⌊N/2⌋1m2≤∫L∞1y2​𝑑y=1L.\displaystyle\sum_{m=L+1}^{\lfloor N/2\rfloor}\frac{1}{m^{2}}\leq\int_{L}^{\infty}\frac{1}{y^{2}}\mathrm{d}y=\frac{1}{L}. (57)
errori≤α​N​‖x‖∞8​L.\displaystyle\mathrm{error}_{i}\leq\frac{\alpha N\|x\|_{\infty}}{8L}. (58)
2​(1−cos⁡ωm)=4​sin2⁡(ωm2)=4​sin2⁡(π​mN)2(1-\cos\omega_{m})=4\sin^{2}\left(\frac{\omega_{m}}{2}\right)=4\sin^{2}\left(\frac{\pi m}{N}\right)
(C+σt2​fα)2≥4​C​σt2​fα(C+\sigma_{t}^{2}f^{\alpha})^{2}\geq 4C\sigma_{t}^{2}f^{\alpha}
sin2⁡(π​mN)≥4​m2N2\sin^{2}\left(\frac{\pi m}{N}\right)\geq\frac{4m^{2}}{N^{2}}

A.3 Theoretical Justification of the Classifier-Based Annotation

The classifier analysis uses balanced source priors and population cross-entropy. A classifier’s mean target probability on auxiliary samples is linked to a divergence between the distributions, with an additional error term for approximate optimization.

Classifier threshold implies distributional closeness.

For an optimal balanced classifier, the density ratio determines its prediction; cross-entropy excess risk then bounds the deviation of an approximate classifier. The source states a LeCam upper bound of twice the threshold tolerance plus twice the square root of excess risk.

The printed LeCam definition and Theorem 3’s coefficient are inconsistent. With LC defined as twice chi-squared divergence from the mixture, direct substitution gives one-half minus LC/4 for the mean classifier output, while the source prints LC/2. The theorem constants and proof are retained unchanged and should not be treated as independently verified guarantees.

Equations 59
ℒCE​(f)=−𝔼c,x​[(1−c)​log⁡(1−f⁡(x))+c​log⁡f⁡(x)],\mathcal{L}_{\mathrm{CE}}(f)=-\mathbb{E}_{c,x}\!\left[(1-c)\log(1-f(x))+c\log f(x)\right],
f∗=arg⁡minf​ℒCE​(f),f∗​(x)=d​Pd​P+d​Q​(x)=12​d​Pd​M​(x).f^{*}=\arg\min_{f}\mathcal{L}_{\mathrm{CE}}(f),\qquad f^{*}(x)=\frac{dP}{dP+dQ}(x)=\frac{1}{2}\frac{dP}{dM}(x). (59)
Theorem 3 (Optimal classifier threshold implies distributional closeness).

Let f∗f^{*} be the optimal classifier in Eq. (59). Then

𝔼Q​f∗​(x)=12−12​LC​(P,Q).\mathbb{E}_{Q}f^{*}(x)=\frac{1}{2}-\frac{1}{2}\mathrm{LC}(P,Q).

In particular, 𝔼Q​f∗​(x)≥12−τ\mathbb{E}_{Q}f^{*}(x)\geq\tfrac{1}{2}-\tau implies LC⁡(P,Q)≤2​τ\mathrm{LC}(P,Q)\leq 2\tau.

𝔼Q​f∗=∫d​P​d​Qd​P+d​Q=12​∫(d​P+d​Q)2−d​P2−d​Q2d​P+d​Q.\mathbb{E}_{Q}f^{*}=\int\frac{dP\,dQ}{dP+dQ}=\frac{1}{2}\int\frac{(dP+dQ)^{2}-dP^{2}-dQ^{2}}{dP+dQ}.
𝔼Q​f∗=1−12​∫d​P2d​P+d​Q−12​∫d​Q2d​P+d​Q.\mathbb{E}_{Q}f^{*}=1-\frac{1}{2}\int\frac{dP^{2}}{dP+dQ}-\frac{1}{2}\int\frac{dQ^{2}}{dP+dQ}.
𝔼Qf∗=12−12χ2(P∥M)−12χ2(Q∥M).\mathbb{E}_{Q}f^{*}=\frac{1}{2}-\frac{1}{2}\chi^{2}(P\|M)-\frac{1}{2}\chi^{2}(Q\|M).
𝔼Q​f∗=12−12​LC​(P,Q).\mathbb{E}_{Q}f^{*}=\frac{1}{2}-\frac{1}{2}\mathrm{LC}(P,Q).
Theorem 4 (Approximate classifier threshold implies distributional closeness).

Suppose fθ:ℝd→[0,1]f_{\theta}:\mathbb{R}^{d}\to[0,1] is an ε\varepsilon-optimal classifier, i.e. ℒCE​(fθ)≤ℒCE​(f∗)+ε\mathcal{L}_{\mathrm{CE}}(f_{\theta})\leq\mathcal{L}_{\mathrm{CE}}(f^{*})+\varepsilon. Then 𝔼Q​fθ​(x)≥12−τ\mathbb{E}_{Q}f_{\theta}(x)\geq\tfrac{1}{2}-\tau implies LC⁡(P,Q)≤2​τ+2​ε\mathrm{LC}(P,Q)\leq 2\tau+2\sqrt{\varepsilon}.

ℒCE(f)=𝔼x∼Md(f∗(x)∥f(x))+const,\mathcal{L}_{\mathrm{CE}}(f)=\mathbb{E}_{x\sim M}\,d(f^{*}(x)\|f(x))+\text{const},
𝔼x∼Q​(fθ​(x)−f∗​(x))2≤ε.\mathbb{E}_{x\sim Q}(f_{\theta}(x)-f^{*}(x))^{2}\leq\varepsilon.
𝔼x∼Q​f∗​(x)≥12−τ−ε.\mathbb{E}_{x\sim Q}f^{*}(x)\geq\frac{1}{2}-\tau-\sqrt{\varepsilon}.
𝔼x∼Q​f∗​(x)≥𝔼x∼Q​fθ​(x)−ε\mathbb{E}_{x\sim Q}f^{*}(x)\geq\mathbb{E}_{x\sim Q}f_{\theta}(x)-\sqrt{\varepsilon}

Appendix B Variance-Preserving vs Variance Exploding Diffusion Processes

Dividing a variance-exploding sample by its time-dependent standard-deviation scale yields the variance-preserving representation with the same signal-to-noise ratio. The density change of variables gives the score scaling, and Tweedie’s identity relates the corresponding optimal denoisers. All three propositions and Equations 60–69 are retained.

Equations 60–69
Xt\displaystyle X_{t} =X0+σVE​(t)​Z,\displaystyle=X_{0}+\sigma_{\mathrm{VE}}(t)\,Z, (60)
Xt\displaystyle X_{t} =1−σVP2​(t)​X0+σVP​(t)​Z,\displaystyle=\sqrt{1-\sigma_{\mathrm{VP}}^{2}(t)}\,X_{0}+\sigma_{\mathrm{VP}}(t)\,Z, (61)
Proposition 1.

The VP and VE processes are equivalent up to a time-dependent rescaling of space.

σVP​(t):=σVE​(t)1+σVE2​(t),c⁡(t):=1+σVE2​(t).\sigma_{\mathrm{VP}}(t):=\frac{\sigma_{\mathrm{VE}}(t)}{\sqrt{1+\sigma_{\mathrm{VE}}^{2}(t)}},\qquad c(t):=\sqrt{1+\sigma_{\mathrm{VE}}^{2}(t)}. (62)
SNRVE​(t)=1σVE2​(t)=1−σVP2​(t)σVP2​(t)=SNRVP​(t),\mathrm{SNR}_{\mathrm{VE}}(t)=\frac{1}{\sigma_{\mathrm{VE}}^{2}(t)}=\frac{1-\sigma_{\mathrm{VP}}^{2}(t)}{\sigma_{\mathrm{VP}}^{2}(t)}=\mathrm{SNR}_{\mathrm{VP}}(t), (63)
X~t:=XtVEc⁡(t)=XtVP.\displaystyle\widetilde{X}_{t}:=\frac{X_{t}^{\mathrm{VE}}}{c(t)}=X_{t}^{\mathrm{VP}}. (64)
Proposition 2.

The score functions of the VP and VE processes are related by:

sVP​(x~t,t)=c⁡(t)⋅sVE​(c⁡(t)​x~t,t).s^{\mathrm{VP}}(\widetilde{x}_{t},t)=c(t)\cdot s^{\mathrm{VE}}\!\left(c(t)\,\widetilde{x}_{t},\,t\right). (65)
Original mathematical display with its complete source layout, enclosures, spacing and notation.
Exact source MathML
sVP​(xtVP,t)\displaystyle s^{\mathrm{VP}}(x_{t}^{\mathrm{VP}},t) =∇xtVP​log​ptVP​(xtVP)\displaystyle=\nabla_{x_{t}^{\mathrm{VP}}}\log p^{\mathrm{VP}}_{t}(x_{t}^{\mathrm{VP}})
=∇xtVP​log​ptVE​(c⁡(t)​xtVP)+∇xtVP​log​c​(t)d0\displaystyle=\nabla_{x_{t}^{\mathrm{VP}}}\log p^{\mathrm{VE}}_{t}(c(t)\,x_{t}^{\mathrm{VP}})+\cancelto{0}{\nabla_{x_{t}^{\mathrm{VP}}}\log c(t)^{d}}
=c⁡(t)⋅sVE​(c⁡(t)​xtVP,t),\displaystyle=c(t)\cdot s^{\mathrm{VE}}\!\left(c(t)\,x_{t}^{\mathrm{VP}},\,t\right), (66)
Proposition 3.

Let hVE∗h^{*}_{\mathrm{VE}} and hVP∗h^{*}_{\mathrm{VP}} denote the optimal denoisers for the VE and VP processes respectively. Then:

hVP∗​(xtVP,t)=hVE∗​(c⁡(t)​xtVP,t).h^{*}_{\mathrm{VP}}(x^{\mathrm{VP}}_{t},t)=h^{*}_{\mathrm{VE}}\!\left(c(t)\,x^{\mathrm{VP}}_{t},\,t\right). (67)
hVE∗​(xtVE,t)\displaystyle h^{*}_{\mathrm{VE}}(x_{t}^{\mathrm{VE}},t) =xtVE+σVE2​(t)​sVE​(xtVE,t),\displaystyle=x_{t}^{\mathrm{VE}}+\sigma_{\mathrm{VE}}^{2}(t)\,s^{\mathrm{VE}}(x_{t}^{\mathrm{VE}},t), (68)
hVP∗​(xtVP,t)\displaystyle h^{*}_{\mathrm{VP}}(x_{t}^{\mathrm{VP}},t) =xtVP+σVP2​(t)​sVP​(xtVP,t)1−σVP2​(t).\displaystyle=\frac{x_{t}^{\mathrm{VP}}+\sigma_{\mathrm{VP}}^{2}(t)\,s^{\mathrm{VP}}(x_{t}^{\mathrm{VP}},t)}{\sqrt{1-\sigma_{\mathrm{VP}}^{2}(t)}}. (69)
xtVE↦xtVE/c⁡(t)=xtVPx^{\mathrm{VE}}_{t}\mapsto x^{\mathrm{VE}}_{t}/c(t)=x^{\mathrm{VP}}_{t}
ptVP​(xtVP)=c​(t)d​ptVE​(c⁡(t)​xtVP)p^{\mathrm{VP}}_{t}(x_{t}^{\mathrm{VP}})=c(t)^{d}\,p^{\mathrm{VE}}_{t}(c(t)\,x_{t}^{\mathrm{VP}})

Appendix C VP Ambient Loss Derivation

The optional Ambient objective targets an intermediate noisy sample without requiring a clean prediction target. A double-Tweedie identity gives affine relationships for either noise prediction or clean-sample prediction; rescaling yields the two variance-preserving losses with the same ideal denoiser minimizer.

At a zero lower cutoff they reduce to ordinary diffusion MSE. Near the lower cutoff, coefficients diverge, so a buffer rejects unstable time/sample combinations. This buffer is sensitive in the pushing ablation; most experiments therefore use the ordinary loss.

Equations 70–82
Proposition 4.

The ϵ\epsilon-prediction VP Ambient loss for t>tmint>t_{\min} is given by

ℒ⁡(θ)=𝔼⁡[‖ϵθ​(xt,t)−σt​(1−σtmin2)σt2−σtmin2​xt+σt​(1−σt2)​(1−σtmin2)σt2−σtmin2​xtmin‖22],\mathcal{L}(\theta)=\mathbb{E}\!\left[\left\lVert\epsilon_{\theta}(x_{t},t)-\frac{\sigma_{t}(1-\sigma_{t_{\min}}^{2})}{\sigma_{t}^{2}-\sigma_{t_{\min}}^{2}}\,x_{t}+\frac{\sigma_{t}\sqrt{(1-\sigma_{t}^{2})(1-\sigma_{t_{\min}}^{2})}}{\sigma_{t}^{2}-\sigma_{t_{\min}}^{2}}\,x_{t_{\min}}\right\rVert_{2}^{2}\right], (70)

and minimized by ϵθ∗=𝔼⁡[Z∣Xt=xt]\epsilon_{\theta}^{*}=\mathbb{E}[Z\mid X_{t}=x_{t}].

𝔼⁡[Z∣Xt=xt]=xt−1−σt2​𝔼​[X0∣Xt=xt]σt.\mathbb{E}[Z\mid X_{t}=x_{t}]=\frac{x_{t}-\sqrt{1-\sigma_{t}^{2}}\,\mathbb{E}[X_{0}\mid X_{t}=x_{t}]}{\sigma_{t}}. (71)
𝔼⁡[X0∣Xt=xt]=σt2​1−σtmin2σt2−σtmin2​𝔼​[Xtmin∣Xt=xt]−σtmin2​1−σt2σt2−σtmin2​xt.\mathbb{E}[X_{0}\mid X_{t}=x_{t}]=\frac{\sigma_{t}^{2}\sqrt{1-\sigma_{t_{\min}}^{2}}}{\sigma_{t}^{2}-\sigma_{t_{\min}}^{2}}\mathbb{E}[X_{t_{\min}}\mid X_{t}=x_{t}]-\frac{\sigma_{t_{\min}}^{2}\sqrt{1-\sigma_{t}^{2}}}{\sigma_{t}^{2}-\sigma_{t_{\min}}^{2}}x_{t}. (72)
𝔼⁡[Xtmin∣Xt=xt]\displaystyle\mathbb{E}[X_{t_{\min}}\mid X_{t}=x_{t}] =αt​𝔼​[Z∣Xt=xt]+βt​xt,\displaystyle=\alpha_{t}\,\mathbb{E}[Z\mid X_{t}=x_{t}]+\beta_{t}\,x_{t}, (73)
αt\displaystyle\alpha_{t} :=−σt2−σtmin2σt​(1−σt2)​(1−σtmin2),\displaystyle:=-\frac{\sigma_{t}^{2}-\sigma_{t_{\min}}^{2}}{\sigma_{t}\sqrt{(1-\sigma_{t}^{2})(1-\sigma_{t_{\min}}^{2})}}, (74)
βt\displaystyle\beta_{t} :=1−σtmin2(1−σt2)​(1−σtmin2).\displaystyle:=\frac{1-\sigma_{t_{\min}}^{2}}{\sqrt{(1-\sigma_{t}^{2})(1-\sigma_{t_{\min}}^{2})}}. (75)
𝔼⁡[∥gθ​(xt,t)−xtmin∥2]\mathbb{E}\!\left[\lVert g_{\theta}(x_{t},t)-x_{t_{\min}}\rVert^{2}\right] (76)
𝔼⁡[‖ϵθ​(xt,t)+βtαt​xt−1αt​xtmin‖22],\mathbb{E}\!\left[\left\lVert\epsilon_{\theta}(x_{t},t)+\frac{\beta_{t}}{\alpha_{t}}x_{t}-\frac{1}{\alpha_{t}}x_{t_{\min}}\right\rVert_{2}^{2}\right], (77)
Proposition 5.

The x0x_{0}-prediction VP Ambient loss for t>tmint>t_{\min} is given by

𝔼⁡[‖hθ​(xt,t)+σtmin2​1−σt2σt2−σtmin2​xt−σt2​1−σtmin2σt2−σtmin2​xtmin‖22],\mathbb{E}\!\left[\left\lVert h_{\theta}(x_{t},t)+\frac{\sigma_{t_{\min}}^{2}\sqrt{1-\sigma_{t}^{2}}}{\sigma_{t}^{2}-\sigma_{t_{\min}}^{2}}x_{t}-\frac{\sigma_{t}^{2}\sqrt{1-\sigma_{t_{\min}}^{2}}}{\sigma_{t}^{2}-\sigma_{t_{\min}}^{2}}x_{t_{\min}}\right\rVert_{2}^{2}\right], (78)

and minimized by hθ∗=𝔼⁡[X0∣Xt=xt]h_{\theta}^{*}=\mathbb{E}[X_{0}\mid X_{t}=x_{t}].

𝔼⁡[Xtmin∣Xt=xt]\displaystyle\mathbb{E}[X_{t_{\min}}\mid X_{t}=x_{t}] =α~t​𝔼​[X0∣Xt=xt]+β~t​xt,\displaystyle=\tilde{\alpha}_{t}\,\mathbb{E}[X_{0}\mid X_{t}=x_{t}]+\tilde{\beta}_{t}\,x_{t}, (79)
α~t\displaystyle\tilde{\alpha}_{t} :=σt2−σtmin2σt2​1−σtmin2,β~t:=σtmin2​1−σt2σt2​1−σtmin2.\displaystyle:=\frac{\sigma_{t}^{2}-\sigma_{t_{\min}}^{2}}{\sigma_{t}^{2}\sqrt{1-\sigma_{t_{\min}}^{2}}},\quad\tilde{\beta}_{t}:=\frac{\sigma_{t_{\min}}^{2}\sqrt{1-\sigma_{t}^{2}}}{\sigma_{t}^{2}\sqrt{1-\sigma_{t_{\min}}^{2}}}. (80)
𝔼⁡[∥g~θ​(xt,t)−xtmin∥2]\mathbb{E}\!\left[\lVert\tilde{g}_{\theta}(x_{t},t)-x_{t_{\min}}\rVert^{2}\right] (81)
𝔼⁡[‖hθ​(xt,t)+β~tα~t​xt−1α~t​xtmin‖22],\mathbb{E}\!\left[\left\lVert h_{\theta}(x_{t},t)+\frac{\tilde{\beta}_{t}}{\tilde{\alpha}_{t}}x_{t}-\frac{1}{\tilde{\alpha}_{t}}x_{t_{\min}}\right\rVert_{2}^{2}\right], (82)
Remark 1.

Both losses reduce to the canonical MSE diffusion loss [54] when tmin=0t_{\min}=0.

Remark 2.

For both the ϵ\epsilon- and x0x_{0}-prediction losses, the rescaling factor, 1/αt21/\alpha_{t}^{2} (resp. 1/α~t21/\tilde{\alpha}_{t}^{2}), diverges as t→tmint\rightarrow t_{\min}. This can cause training instability. Following [19], we introduce a buffer BB: a sample from 𝒟q\mathcal{D}_{q} is only used at diffusion times satisfying t>tmint>t_{\min} and 1/αt2<B1/\alpha_{t}^{2}<B (resp. 1/α~t2<B1/\tilde{\alpha}_{t}^{2}<B). The effect of BB is studied in Figure 1919(e).

gθ​(xt,t):=αt​ϵθ​(xt,t)+βt​xtg_{\theta}(x_{t},t):=\alpha_{t}\,\epsilon_{\theta}(x_{t},t)+\beta_{t}\,x_{t}
g~θ​(xt,t):=α~t​hθ​(xt,t)+β~t​xt\tilde{g}_{\theta}(x_{t},t):=\tilde{\alpha}_{t}\,h_{\theta}(x_{t},t)+\tilde{\beta}_{t}\,x_{t}
g~θ∗=𝔼⁡[Xtmin∣Xt=xt]\tilde{g}_{\theta}^{*}=\mathbb{E}[X_{t_{\min}}\mid X_{t}=x_{t}]

Appendix D Characterization of Robot Action Data and Action Diffusion

PSD measurements normalize action ranges, use 100-step chunks and average per-dimension periodograms. OXE uses 2nd/98th percentile scaling and is converted to 10 Hz by downsampling faster sources and speeding up slower ones; DC is omitted. These preprocessing choices matter when interpreting the observed spectra.

Table 4 retains seven dataset descriptions and fitted slopes. The source says nine-dimensional rotation entries lie in [0,1], although rotation-matrix entries generally span [−1,1]; the actual stated choice is to leave those dimensions unnormalized.

A classifier trained on 5,000 maze trajectories predicts which of 20 convex regions a noisy path visits, excluding start/goal conditioning. It reaches 95% accuracy by noise .878, supporting early recovery of global path information. Equation 83 separately measures normalized Jacobian sensitivity of the eighth action to all sixteen input actions; normalization removes absolute scale differences across noise levels.

Complete original figure with every panel, label and legend.
Figure 11: A visualization of the average PSD for a diverse range of robot datasets; a spectral power law exists in all of them. Table 4 characterizes the datasets shown in (a)-(f) as well as the OXE dataset shown in Figure 3(a).
Selectable original panel labels

(a) 𝒟p\mathcal{D}_{p} from Section 7.1

(b) 𝒟p\mathcal{D}_{p} from Appendix G.2

(c) 𝒟p\mathcal{D}_{p} from Section 7.3

(d) 𝒟q\mathcal{D}_{q} from Section 7.3

(e) 𝒟p\mathcal{D}_{p} from Section 7.4

(f) 𝒟p\mathcal{D}_{p} from Section 8

Table 4: Characteristics of the datasets analyzed via power spectral density (PSD) in Figures 3(a) and 11. Despite their differences, all datasets exhibit a spectral power law.
Fig. Dataset Source Generation Action Space Action Type Rotation Rep. α¯†\bar{\alpha}^{\dagger}
Fig. 3(a) COXE [1] 𝒟q\mathcal{D}_{q} (Sec. 8) Mixed Task space Delta Axis-angle -1.47
Fig. 11(a) GCS [65] 𝒟p\mathcal{D}_{p} (Sec. 7.1) Algorithmic xx-yy Absolute N/A -1.82
Fig. 11(b) TrajOpt. 𝒟p\mathcal{D}_{p} (App. G.2) Algorithmic Joint space Absolute N/A -1.82
Fig. 11(c) Planar pushing [12] 𝒟p\mathcal{D}_{p} (Sec. 7.3) Human teleop xx-yy Absolute N/A -2.01
Fig. 11(d) Graesdal et al. [72] 𝒟q\mathcal{D}_{q} (Sec. 7.3) Algorithmic xx-yy Absolute N/A -1.88
Fig. 11(e) Block sorting 𝒟p\mathcal{D}_{p} (Sec. 7.4) Human teleop Task space Absolute ℝ9\mathbb{R}^{9} -2.27
Fig. 11(f) Table cleaning 𝒟p\mathcal{D}_{p} (Sec. 8) Human teleop Task space Absolute ℝ9\mathbb{R}^{9} -2.17

†\dagger α¯\bar{\alpha}: mean fitted exponent across action dimensions, where S⁡(f)∝fαS(f)\propto f^{\alpha}

Complete original figure with every panel, label and legend.
Figure 12: The backward process commits to a global plan at high noise. By σt=0.878\sigma_{t}=0.878, the global plan in A0A_{0} can already be recovered from AtA_{t} with 95% accuracy. Accuracy plateaus afterward. This indicates that the start of the backward process performs global planning. Figure 2 shows that the remainder of the backward process is devoted to local trajectory refinement.
Selectable original panel labels

(a) Convex decomposition of the collision-free space.

(b) Visualization of the discrete decisions in GCS.

(c) Prediction accuracy for the plan at different noise levels.

Equation 83
St​(i):=𝔼(O,A0)∼𝒟​[‖∂hθ(8)​(At,O,t)∂At(i)‖F],i∈[0,15].S_{t}(i):=\mathbb{E}_{(O,A_{0})\sim\mathcal{D}}\!\left[\left\lVert\frac{\partial h_{\theta}^{(8)}(A_{t},O,t)}{\partial A_{t}^{(i)}}\right\rVert_{F}\right],\quad i\in[0,15]. (83)

Appendix E Pseudocode and Possible Extensions

The implementation appendix retains the complete training and loss pseudocode, followed by an untested classifier rejection-sampling extension.

E.1 Ambient Diffusion Policy Pseudocode

Algorithm 1 samples times before admissible examples. Algorithms 2 and 3 implement variance-exploding and variance-preserving forward processes, using ordinary denoising at low noise and optionally the Ambient objective above the lower cutoff. The source’s first listing line is a Require declaration in the PDF, not an executed numbered operation.

Only the maze experiments use the Ambient loss in the reported main configurations. Testing it on OXE found no significant difference, so the ordinary loss was retained for simplicity.

Algorithm 1 Ambient Diffusion Policy: Training
Complete original pseudocode with every line and its original nesting.
Selectable pseudocode
Require: 𝒟p\mathcal{D}_{p}, 𝒟q\mathcal{D}_{q} with (tmin(i),tmax(i))(t_{\min}^{(i)},t_{\max}^{(i)}), ℒ∈{ℒambient,ℒdenoising}\mathcal{L}\in\{\mathcal{L}_{\text{ambient}},\mathcal{L}_{\text{denoising}}\}, noise schedule σ⁡(t)\sigma(t), max time TT, batch size BB, policy parameters θ\theta
1: repeat
2:   Sample t(i)∼i.i.d.𝒰⁡[0,T]t^{(i)}\stackrel{{\scriptstyle\text{i.i.d.}}}{{\sim}}\mathcal{U}[0,T] for i=1,…,Bi=1,\ldots,B
3:   Sample batch {(O(i),A0(i))}i=1B\{(O^{(i)},A_{0}^{(i)})\}_{i=1}^{B} s.t. t(i)∈[0,tmax(i))∪(tmin(i),T]t^{(i)}\in[0,t_{\max}^{(i)})\cup(t_{\min}^{(i)},T]
4:   ℓ←1B​∑i=1BComputeLoss​(ℒ,A0(i),O(i),t(i),tmax(i),tmin(i))\ell\leftarrow\tfrac{1}{B}\sum_{i=1}^{B}\textsc{ComputeLoss}\left(\mathcal{L},A_{0}^{(i)},O^{(i)},t^{(i)},t_{\max}^{(i)},t_{\min}^{(i)}\right)
5:   Update θ\theta on ℓ\ell
6: until converged
7: return πθ\pi_{\theta}
Algorithm 2 ComputeLoss (variance-exploding)
Complete original pseudocode with every line and its original nesting.
Selectable pseudocode
Require: ℒ\mathcal{L}, A0A_{0}, OO, tt, tmaxt_{\max}, tmint_{\min}
1: if t∈[0,tmax)t\in[0,t_{\max}) or ℒ=ℒdenoising\mathcal{L}=\mathcal{L}_{\text{denoising}} then
2:   At←A0+σt​ZA_{t}\leftarrow A_{0}+\sigma_{t}\,Z,  Z∼𝒩⁡(0,I)Z\sim\mathcal{N}(0,I)
3:   return ℒdenoising​(At,A0,O,t)\mathcal{L}_{\text{denoising}}(A_{t},A_{0},O,t) ⊳\triangleright standard DDPM loss [54]
4: else⊳\triangleright t∈(tmin,T]t\in(t_{\min},T] and ℒ=ℒambientVE\mathcal{L}=\mathcal{L}_{\text{ambient}}^{\text{VE}}
5:   Sample Z1,Z2∼i.i.d.𝒩⁡(0,I)Z_{1},Z_{2}\stackrel{{\scriptstyle\text{i.i.d.}}}{{\sim}}\mathcal{N}(0,I)
6:   Atmin←A0+σtmin​Z1A_{t_{\min}}\leftarrow A_{0}+\sigma_{t_{\min}}\,Z_{1}
7:   At←Atmin+σt2−σtmin2​Z2A_{t}\leftarrow A_{t_{\min}}+\sqrt{\sigma_{t}^{2}-\sigma_{t_{\min}}^{2}}\ Z_{2}
8:   return ℒambientVE​(At,Atmin,O,t)\mathcal{L}_{\text{ambient}}^{\text{VE}}(A_{t},A_{t_{\min}},O,t) ⊳\triangleright VE Ambient loss [24]
9: end if
Algorithm 3 ComputeLoss (variance-preserving)
Complete original pseudocode with every line and its original nesting.
Selectable pseudocode
Require: ℒ\mathcal{L}, A0A_{0}, OO, tt, tmaxt_{\max}, tmint_{\min}
1: if t∈[0,tmax)t\in[0,t_{\max}) or ℒ=ℒdenoising\mathcal{L}=\mathcal{L}_{\text{denoising}} then
2:   At←1−σt2​A0+σt​ZA_{t}\leftarrow\sqrt{1-\sigma_{t}^{2}}A_{0}+\sigma_{t}Z,  Z∼𝒩⁡(0,I)Z\sim\mathcal{N}(0,I)
3:   return ℒdenoising​(At,A0,O,t)\mathcal{L}_{\text{denoising}}(A_{t},A_{0},O,t) ⊳\triangleright standard DDPM loss [54]
4: else⊳\triangleright t∈(tmin,T]t\in(t_{\min},T] and ℒ=ℒambientVP\mathcal{L}=\mathcal{L}_{\text{ambient}}^{\text{VP}}
5:   Sample Z1,Z2∼i.i.d.𝒩⁡(0,I)Z_{1},Z_{2}\stackrel{{\scriptstyle\text{i.i.d.}}}{{\sim}}\mathcal{N}(0,I)
6:   Atmin←1−σtmin2​A0+σtmin​Z1A_{t_{\min}}\leftarrow\sqrt{1-\sigma_{t_{\min}}^{2}}A_{0}+\sigma_{t_{\min}}Z_{1}
7:   At←1−σt21−σtmin2​Atmin+σt2−σtmin21−σtmin2​Z2A_{t}\leftarrow\sqrt{\frac{1-\sigma_{t}^{2}}{1-\sigma_{t_{\min}}^{2}}}A_{t_{\min}}+\sqrt{\frac{\sigma_{t}^{2}-\sigma_{t_{\min}}^{2}}{1-\sigma_{t_{\min}}^{2}}}Z_{2}
8:   return ℒambientVP​(At,Atmin,O,t)\mathcal{L}_{\text{ambient}}^{\text{VP}}(A_{t},A_{t_{\min}},O,t) ⊳\triangleright Eq. (70) (ϵ\epsilon-pred) or Eq. (78) (x0x_{0}-pred)
9: end if
Complete original figure with every panel, label and legend.
Figure 13: Average classifier output for suboptimal samples in the 2D maze dataset (Section 7.1) at different diffusion times. As tt increases, ptp_{t} and qtq_{t} become harder to distinguish. The dotted line is the threshold 0.5−τ0.5-\tau, which is crossed for tmin=18t_{\min}=18.

E.2 Diffusion Policy Implementation Details

Unless otherwise specified, a U-Net predicts sixteen actions at 10 Hz from two observations, executes eight and replans. Image-conditioned policies use ResNet18. Training uses a cosine learning-rate schedule, 100 cosine-scheduled diffusion steps and exponential moving averages; inference uses ten DDIM steps.

Predictions are linearly interpolated and passed through a first-order low-pass filter before execution. The neural motion planner instead predicts a complete 100-waypoint plan, as specified in its task appendix.

E.3 Future Direction: Classifier-Based Rejection Sampling

Algorithm 4 samples an equal target/auxiliary mixture, adds noise and accepts according to the classifier’s target probability. With uniformly small population cross-entropy error at each noise level, the proof bounds rejection-sampled distribution error by an order-square-root term.

The proof controls acceptance probability around one half and requires sufficiently small error for the explicit constant. This variant has no experimental validation in the paper; it is not the sampler used for the reported main results.

Algorithm 4 Rejection sampling via classifier
Complete original pseudocode with every line and its original nesting.
Selectable pseudocode
Require: Noisy classifier fθ​(x,σ)f_{\theta}(x,\sigma), mixture distribution M=12​P+12​QM=\tfrac{1}{2}P+\tfrac{1}{2}Q, noise distribution ρσ\rho_{\sigma}
1: Sample σ∼ρσ\sigma\sim\rho_{\sigma}
2: while True do
3:   Sample X∼MX\sim M
4:   Xσ←AddNoise​(X,σ)X_{\sigma}\leftarrow\textsc{AddNoise}(X,\sigma)
5:   Sample c∼Bern⁡(fθ​(Xσ,σ))c\sim\mathrm{Bern}(f_{\theta}(X_{\sigma},\sigma))
6:   if c=1c=1 then
7:    return (σ,Xσ)(\sigma,X_{\sigma})
8:   end if
9: end while
Equations
Proposition 6 (Rejection sampling guarantee).

Suppose fθf_{\theta} is uniformly ε\varepsilon-optimal across σ\sigma:

ℒCE​(f∣σ)≤ℒCE​(f∗∣σ)+ε,∀σ.\mathcal{L}_{\mathrm{CE}}(f\mid\sigma)\leq\mathcal{L}_{\mathrm{CE}}(f^{*}\mid\sigma)+\varepsilon,\qquad\forall\sigma.

For any σ≥0\sigma\geq 0, let PσP_{\sigma} be the noisy distribution of X∼PX\sim P, and let P^σ\hat{P}_{\sigma} be the distribution of XσX_{\sigma} output by Algorithm 4. Then

dTV​(P^σ,Pσ)≤O⁡(ε).\mathrm{d}_{\mathrm{TV}}(\hat{P}_{\sigma},P_{\sigma})\leq O(\sqrt{\varepsilon}).
p^​(x)=m⁡(x)​fθ​(x)Z,Z=𝔼X∼M​[fθ​(X)],\hat{p}(x)=\frac{m(x)f_{\theta}(x)}{Z},\quad Z=\mathbb{E}_{X\sim M}[f_{\theta}(X)],
δ≜𝔼X∼M​|fθ​(X)−f∗​(X)|≤ε/2.\delta\triangleq\mathbb{E}_{X\sim M}|f_{\theta}(X)-f^{*}(X)|\leq\sqrt{\varepsilon/2}.
dTV​(P^,P)\displaystyle\mathrm{d}_{\mathrm{TV}}(\hat{P},P) =12​∫|m⁡(x)​fθ​(x)Z−2​m​(x)​f∗​(x)|​𝑑x\displaystyle=\frac{1}{2}\int\left|\frac{m(x)f_{\theta}(x)}{Z}-2m(x)f^{*}(x)\right|dx
=𝔼X∼M​|fθ​(X)−2​Z​f∗​(X)2​Z|.\displaystyle=\mathbb{E}_{X\sim M}\left|\frac{f_{\theta}(X)-2Zf^{*}(X)}{2Z}\right|.
|fθ​(X)−2​Z​f∗​(X)2​Z|≤|fθ​(X)−f∗​(X)|2​Z+f∗​(X)2​Z​|1−2​Z|.\left|\frac{f_{\theta}(X)-2Zf^{*}(X)}{2Z}\right|\leq\frac{|f_{\theta}(X)-f^{*}(X)|}{2Z}+\frac{f^{*}(X)}{2Z}|1-2Z|.
dTV​(P^,P)≤δ2​Z+2​δ4​Z=δZ≤2​δ1−2​δ.\mathrm{d}_{\mathrm{TV}}(\hat{P},P)\leq\frac{\delta}{2Z}+\frac{2\delta}{4Z}=\frac{\delta}{Z}\leq\frac{2\delta}{1-2\delta}.
dTV​(P^,P)≤6​ε/2=O⁡(ε).\mathrm{d}_{\mathrm{TV}}(\hat{P},P)\leq 6\sqrt{\varepsilon/2}=O(\sqrt{\varepsilon}).
𝔼X∼M​[(f∗​(X)−fθ​(X))2]≤ε/2\mathbb{E}_{X\sim M}[(f^{*}(X)-f_{\theta}(X))^{2}]\leq\varepsilon/2

Appendix F Experiments Overview

Tables 5–6 retain every dataset size, observation/action specification, cutoff, mixture weight, prediction target and trial count. Binary outcomes use 95% Clopper–Pearson intervals; other bars are standard errors. Five to ten checkpoints are saved per run and the best-performing checkpoint is reported.

These intervals describe evaluation outcomes rather than independent training-seed uncertainty. Hyperparameters are swept for controlled baselines; OXE uses fixed .5 mixture weighting and manually selected low-noise cutoffs without a hardware sweep.

Table 5: Experimental configurations for the four generality experiments. Parameters vary for the ablations.
Maze
Neural
Motion Planning
Sim-and-Real
Co-training
Block
Sorting
Dataset Details 𝒟p\mathcal{D}_{p} GCS [65]
Trajectory
Optimization [83]
Joystick
Teleop [12]
VR Teleop
𝒟q\mathcal{D}_{q} RRT [66] BiRRT [84]
Contact-Rich
Planner [72]
VR Teleop
|𝒟p||\mathcal{D}_{p}| (Num. Demos.) 50 100,000 50 50
|𝒟q||\mathcal{D}_{q}| (Num. Demos.) 5000 1,000,000 2000 200
Nature of shift
Low-quality/
noisy trajectories
Low-quality/
noisy trajectories
Sim-to-real
gap
Task
mismatch
Policy I/O Observation space Proprio Proprio
EE position
Scene camera
Wrist camera
EE pose (xyz, R9)
Gripper state
Scene camera
Blocks camera
Wrist camera
Action space xx-yy position
Joint angles
ℝ7\mathbb{R}^{7}
xx-yy pusher
position
EE command
(xyz, R9, gripper)
Parameters σtmin∗∈[0,1]\sigma_{t_{\min}^{*}}\in[0,1]
0.074
(sweep)
0.025
(sweep)
varies
see Table 2
N/A
(sweep)
σtmax∗∈[0,1]\sigma_{t_{\max}^{*}}\in[0,1] N/A N/A
0
(sweep)
0.46
(sweep)
α∗∈[0,1]\alpha^{*}\in[0,1] for co-training
0.019
(sweep)
0.091
(sweep)
0.25
(sweep)
0.9
(sweep)
Loss Loss prediction target x0x_{0} ϵ\epsilon x0x_{0} ϵ\epsilon
Loss function ℒambient\mathcal{L}_{\text{ambient}} ℒdenoising\mathcal{L}_{\text{denoising}} ℒdenoising\mathcal{L}_{\text{denoising}} ℒdenoising\mathcal{L}_{\text{denoising}}

Eval.

Num. trials 1,000 1,000 200
200
( 800 blocks)
Table 6: Experimental configurations for the two OXE scaling experiments. Parameters vary for the ablations.
Table
Cleaning
Tower
Building
Dataset Details 𝒟p\mathcal{D}_{p} VR Teleop VR Teleop
𝒟q\mathcal{D}_{q}
[1]
[1]
|𝒟p||\mathcal{D}_{p}| (Num. Demos.) 50 & 150 35
|𝒟q||\mathcal{D}_{q}| (Num. Demos.)
1,279,774 (MS++)
1,402,606 (COXE)
1,402,606 (COXE)
Nature of shift Unstructured Unstructured
Policy I/O Observation space
EE pose (xyz, R9)
Gripper state
Table camera
Drawer camera
2x Wrist camera
EE pose (xyz, R9)
Gripper state
Table camera
Tower camera
2x Wrist camera
Action space
Delta pose
(xyz, axis-angle,
gripper)
Delta pose
(xyz, axis-angle,
gripper)
Parameters σtmin∗∈[0,1]\sigma_{t_{\min}^{*}}\in[0,1]
See Figure 21
(classifier)
See Figure 22
(classifier)
σtmax∗∈[0,1]\sigma_{t_{\max}^{*}}\in[0,1]
0.167
(no sweep)
0.09
(no sweep)
α∗∈[0,1]\alpha^{*}\in[0,1] for co-training
0.5
(no sweep)
0.5
(no sweep)
Loss Loss prediction target ϵ\epsilon ϵ\epsilon
Loss function ℒdenoising\mathcal{L}_{\text{denoising}} ℒdenoising\mathcal{L}_{\text{denoising}}

Eval.

Num. trials
20
(60 objects)
20

Appendix G Controlled Experiments

The controlled-task appendix specifies data generation and evaluation, including conditional smoothness statistics, planning tolerances and the distinction between simulated transfer and hardware experiments.

G.1 Noisy Trajectories: 2D Maze

The 5×5 m maze uses third-order Bézier GCS paths with first-order continuity and up to ten rounded candidates. Both datasets impose 1 m/s componentwise velocity limits and .1 m obstacle padding. Success means reaching within .15 m of the goal without collision in 30 seconds.

Smoothness is time-averaged squared acceleration of a fitted cubic spline, reported only for successful rollouts. Figure 14 retains matched start/goal examples; Figure 15 shows the coverage loss as higher cutoffs admit less RRT data.

Equation 84
Smoothness=1T​∫0T∥a⁡(t)∥22​𝑑t,\mathrm{Smoothness}=\tfrac{1}{T}\!\!\int_{0}^{T}\!\!\lVert a(t)\rVert_{2}^{2}\,dt, (84)
Complete original figure with every panel, label and legend.
Figure 14: Policy rollouts from the same start and goal pairs for different training algorithms. a) The policy trained with data filtering (i.e. GCS-only) has not seen enough data to learn the maze. Its rollouts frequently collide with obstacles. b) The policy trained with RRT-only has memorized the maze, but exhibits non-smoothness. c) The co-trained policy’s behavior is similar to the RRT-only policy. d) The Ambient Diffusion Policy is smooth and understands the structure of the maze.
Complete original figure with every panel, label and legend.
Figure 15: Maze success rate vs σtmin.\sigma_{t_{\min}}. For small σtmin\sigma_{t_{\min}}, the Ambient Policies are smooth and achieve high success rate (see Figure 6(a)). As σtmin\sigma_{t_{\min}} increases, the RRT data is used less during training, which reduces success rate.

G.2 Noisy Trajectories: Neural Motion Planning

Start/goal configurations come equally from shelves, bins, near objects and free space. Inverse kinematics uses 1 cm clearance; BiRRT gives auxiliary paths. Target paths add shortcutting and B-spline optimization with 200 collision-constraint samples, followed by finer collision filtering. All paths become 100 equally timed waypoints.

The optimizer constrains squared joint-speed cost, not the reported squared-acceleration metric. Evaluation tolerates 2 cm penetration and .2-radian endpoint error. Best-of-100 selects the least-colliding predicted plan, raising success to 91.9% for Ambient and 91.3% for co-training; their smoothness is 20.2 versus 32.6. This is a different computation budget from single-plan evaluation.

Table 7: Comparison between the 2D maze and 7-DoF neural motion planning experiments.
2D Maze 7-DoF NMP
𝒟p\mathcal{D}_{p} Source GCS [65] Trajectory Optimizer [83]
𝒟q\mathcal{D}_{q} Source RRT [66] BiRRT [84]
|𝒟p||\mathcal{D}_{p}| 50 100,000
|𝒟q||\mathcal{D}_{q}| 5,000 1,000,000
Output Type Action Chunk Full Plan
Rollouts Closed-Loop Open-Loop
Penetration Threshold 0cm 2cm
Output Space 2D (x,y)(x,y) 7D (Config. Space)
Start & End Goals Uniformly Random
Around points of interest
(See Appendix G.2)
Equations 85–86
findq∈ℝ7s.t.FKpos⁡(q)=p,SDF⁡(q)≥0.01,\begin{array}[]{rl}\operatorname{find}&q\in\mathbb{R}^{7}\\ \mathrm{s.t.}&\operatorname{FK}_{\mathrm{pos}}(q)=p,\\ &\operatorname{SDF}(q)\geq 0.01,\end{array} (85)
minq⁡(s)∫01‖q˙​(s)‖2​𝑑ss.t.q(0)=q0,q(1)=q1,SDF⁡(q⁡(s))≥0,∀s∈[0,1]\begin{array}[]{rl}\min_{q(s)}&\int_{0}^{1}||\dot{q}(s)||^{2}\,\mathrm{d}s\\ \mathrm{s.t.}&q(0)=q_{0},q(1)=q_{1},\\ &\operatorname{SDF}(q(s))\geq 0,\;\forall s\in[0,1]\end{array} (86)
Training Algo. Success Rate ↑\uparrow
Smoothness ↓\downarrow
(Eq. (84))
Data Filtering 73.9−2.8+2.773.9^{+2.7}_{-2.8} % 3.7\mathbf{3.7}
Co-train
(α∗=0.091\alpha^{*}=0.091)
91.3−1.9+1.791.3^{+1.7}_{-1.9} % 32.632.6
Ambient
(σtmin∗=0.025\sigma_{t_{\min}^{*}}=0.025)
91.9−1.9+1.6\mathbf{91.9^{+1.6}_{-1.9}} % 20.220.2
Table 8: Neural Motion Planning Best-of-100 Results (1000 trials). Ambient retains the highest success rate while achieving lower squared acceleration than the co-trained baseline.
Complete original figure with every panel, label and legend.
Figure 16: Figure 6(b) repeated with best-of-100 evaluation. The overall trends are the same, but the success rates are higher and the gap between the two Pareto frontiers is smaller.
Complete original figure with every panel, label and legend.
Figure 17: Neural motion planning σtmin\sigma_{t_{\min}} sweep. (17(a)) For small σtmin\sigma_{t_{\min}}, the policies are smoother, which results in fewer accidental collisions and increases success rate. For large σtmin\sigma_{t_{\min}}, the RRT data becomes under-utilized, which decreases the success rate. (17(b)) Increasing σtmin\sigma_{t_{\min}} improves smoothness.
Selectable original panel labels

(a) Success rate vs σtmin\sigma_{t_{\min}}.

(b) Smoothness metric vs σtmin\sigma_{t_{\min}}.

G.3 Sim-and-Real Co-training: Planar Pushing

Both source and target are simulators. Their differences include lighting/colors/camera/shadows, quasistatic versus Drake physics and planner versus joystick actions. Success requires returning the pusher to its default position with the object within 1.5 cm and 3.5° of the goal in 75 seconds.

Eight ablations vary lower and upper cutoffs, classifier duration/threshold, Ambient-loss buffer, annotation granularity and auxiliary scale at two target-set sizes. Per-chunk labels perform best among the main variants; random chunk partitions with 10,000 groups are strongest in the partition study.

Sufficiently trained classifiers are relatively insensitive to checkpoint and threshold in this task. The optional loss is sensitive to its buffer. Auxiliary scaling generally favors Ambient, with an explicit exception at 4,000 auxiliary demonstrations and 50 target demonstrations. Locality does not help when the local contact dynamics differ.

Table 9: Sim-to-target gap along three axes. The sim-to-target gap in the environments from [12] was designed to mimic the sim-to-real gap. See [12] for visualizations.
Gap Simulation Target Sim
Visual
White spectrum light
Different object colors
Different camera pose
No shadows
Warm spectrum light
Different object colors
Different camera pose
Shadows
Physics Quasistatic Drake physics [90]
Action Motion planning Human teleop
Equation
tmin=inf{t:𝔼(O,A0)∼Pi,At|A0​[cϕ∗​(At,t)]>0.5−τ}.t_{\min}=\inf\{t:\mathbb{E}_{(O,A_{0})\sim P_{i},A_{t}|A_{0}}[c_{\phi^{*}}(A_{t},t)]>0.5-\tau\}.
Complete original figure with every panel, label and legend.
Figure 18: Distribution of tmint_{\min} labels. This histogram shows the distribution of tmin∈[0,100]t_{\min}\in[0,100] assigned to each action chunk by the classifier for the sim-and-real experiments.
Complete original figure with every panel, label and legend.
Figure 19: Sim-and-real co-training ablations. Eight ablations covering the key design decisions of Ambient Diffusion Policy. See Appendix G.3 for the per-panel analysis.
Selectable original panel labels

(a) Performance vs σtmin\sigma_{t_{\min}} (per dataset).

(b) Performance vs σtmax\sigma_{t_{\max}}.

(c) Sensitivity to classifier training duration.

(d) Sensitivity to classifier threshold τ\tau.

(e) Sensitivity to the Ambient loss buffer (Appendix C).

(f) Performance vs granularity of tmint_{\min} labels.

(g) Performance vs |𝒟q||\mathcal{D}_{q}| scale (|𝒟p|=10|\mathcal{D}_{p}|=10).

(h) Performance vs |𝒟q||\mathcal{D}_{q}| scale (|𝒟p|=50|\mathcal{D}_{p}|=50).

G.4 Task Mismatch: Block Sorting

VR teleoperation collects both sorting datasets. Each of 200 trials contains one to three blocks of each color and allows twenty seconds per block, for roughly 800 evaluated blocks. Grasp attempts and time per block improve as the upper cutoff admits more low-noise practice.

With a one-hot task-direction input, co-training reaches 90.3% correct placements and Ambient 92.8%. A separate unconditioned policy using only wrong-task data at low noise still reaches 91.9%, with 97.9% conditional logic and 93.8% placement success. Task conditioning and time-dependent sampling are complementary here.

Complete original figure with every panel, label and legend.
Figure 20: Additional motion-level metrics for the block sorting experiment (Section 7.4). (20(a)) Grasp attempts per block improves as σtmax\sigma_{t_{\max}} increases but plateaus after σtmax∗=0.46\sigma_{t_{\max}^{*}}=0.46. (20(b)) Time per block improves as σtmax\sigma_{t_{\max}} increases.
Selectable original panel labels

(a) Grasp attempts per block vs σtmax\sigma_{t_{\max}}.

(b) Time per block vs σtmax\sigma_{t_{\max}}.

Table 10: Block sorting ablations (200 trials, ∼\sim800 blocks). (a) Ablation 1 (Task-Conditioned): Ambient outperforms co-training. (b) Ablation 2: the Ambient Diffusion Policy performs well despite training on no logically correct samples for t∈[0,tmax∗)t\in[0,t_{\max}^{*}).
(a) Ablation 1: Task-Conditioned (b) Ablation 2
Training Algo. Co-train Locality Locality
(# Blocks: ∼\sim800) (α∗=0.5\alpha^{*}=0.5) (σtmax∗=0.89\sigma_{t_{\max}^{*}}=0.89)
𝒟q\mathcal{D}_{q} only below
σtmax∗=0.24\sigma_{t_{\max}^{*}}=0.24
Logic ↑\uparrow 98.6−1.1+0.7\mathbf{98.6^{+0.7}_{-1.1}} % 98.5−1.1+0.798.5^{+0.7}_{-1.1} % 97.9−1.3+0.997.9^{+0.9}_{-1.3} %
Motion ↑\uparrow 91.5−2.2+1.891.5^{+1.8}_{-2.2} % 94.2−1.8+1.5\mathbf{94.2^{+1.5}_{-1.8}} % 93.8−1.9+1.693.8^{+1.6}_{-1.9} %
Success Rate ↑\uparrow 90.3−2.3+2.090.3^{+2.0}_{-2.3} % 92.8−2.0+1.7\mathbf{92.8^{+1.7}_{-2.0}} % 91.9−2.1+1.891.9^{+1.8}_{-2.1} %

Appendix H Scaling Experiments: Open X-Embodiment

The OXE appendix records curation, weighting, annotation, hardware evaluation and additional behavior measurements.

H.1 OXE Experiments

Custom OXE contains 48 accessible single-arm/mobile-manipulator datasets with compatible position/velocity actions, known frequency, proprioception and an RGB camera. It includes all 27 Magic Soup++ datasets and totals 1,402,606 demonstrations versus 1,279,774. This is an inclusive compatible subset, not all OXE data.

Weights reuse OpenVLA values where available; new-source weights mostly follow size with the two largest halved. Total target/auxiliary weight is .5/.5 before time-dependent normalization. Forty-eight binary classifiers label one lower cutoff per source; a single source-conditioned classifier is suggested but not used.

Upper cutoffs are fixed at diffusion step ten for cleaning and five for towers, without a sweep. A classifier-based upper-cutoff method would assign zero to nearly all sources, so the reported locality settings are manual choices.

Cleaning trains on 31 objects and evaluates three unseen objects per trial for ninety seconds. Opening/closing the drawer each earns .1 and each cleaned object .8/3. A bug supplies 128×128 inference images upscaled to 224×224, whereas training used 640×320 images resized to 224×224; all cleaning comparisons retain this mismatch.

Tower trials count correctly colored/oriented blocks until collapse or human intervention. Both tasks use twenty trials per policy; the camera bug was fixed for towers.

Complete original figure with every panel, label and legend.
Figure 21: Table Cleaning: Dataset re-weighting and tmint_{\min} visualization. These plots show the sampling ratio for 𝒟p\mathcal{D}_{p} (shown in blue) and each dataset in COXE at different diffusion times for the table cleaning task. Left: Co-training uses the same sampling ratio at all diffusion times. Right: The sampling ratio for each dataset varies at each diffusion time. Notably, at intermediate diffusion times, only data from 𝒟p\mathcal{D}_{p} can be used (the “Donut paradox” from Daras et al. [19]). The histogram shows the distribution of tmint_{\min} values assigned by the classifiers to each dataset in COXE.
Complete original figure with every panel, label and legend.
Figure 22: Tower Building: Dataset re-weighting and tmint_{\min} visualization. See Figure 21 for the equivalent plot descriptions.
Table 11: Table cleaning evaluation rubric. The maximum score per trial is 1.0.
Subtask Points
Open drawer 0.1
Clean object (per object, 3 total) 0.8/3≈0.2670.8/3\approx 0.267
Close drawer (after table cleaned) 0.1
Max score per trial 1.0

H.2 OXE Ablations

The additional OXE analysis examines repeated grasping and object-cleaning time, alongside the main weighting ablation and the finetuning appendix.

H.2.1 Qualitative Performance

Long cleaning timeouts can conceal repeated failed grasps. Across ten objects, Magic Soup++ Ambient reduces mean attempts from 2.4 to 1.6 and mean time from 24.9 to 14.8 seconds. Custom OXE reduces attempts from 2.1 to 1.7 and time from 22.1 to 16.1 seconds.

The tables report standard errors across ten per-object values. The custom-OXE attempt reduction is about 19%, smaller than the broad 25–40% prose claim; the exact object-level values remain visible.

Table 12: Average grasp attempts per object. The error bars on the Average row report the standard error of the mean across per-object values in each column (n=10n=10 objects).
MS++ COXE
Object Co-train
Ambient +
Locality
Co-train
Ambient +
Locality
Apple 2 1 2 2
Figurine 4 2 2 2
Eraser 4 2 2 1
Play-Doh 2 2 2 2
Robot Gripper 3 2 2 2
Yarn 2 2 2 2
Camera Mount 1 1 1 2
3D Printed Object 2 2 3 1
Mustard 3 1 2 2
Garbage Bag Box 1 1 3 1
Average 2.4±0.32.4\pm 0.3 1.6±0.21.6\pm 0.2 2.1±0.22.1\pm 0.2 1.7±0.21.7\pm 0.2
Table 13: Average cleaning time per object. The error bars on the Average row report the standard error of the mean across per-object values in each column (n=10n=10 objects).
MS++ COXE
Object Co-train
Ambient +
Locality
Co-train
Ambient +
Locality
Apple 21s 9s 30s 19s
Figurine 44s 20s 30s 20s
Eraser 39s 16s 18s 8s
Play-Doh 20s 20s 19s 20s
Robot Gripper 31s 19s 19s 20s
Yarn 22s 19s 20s 20s
Camera Mount 12s 8s 8s 20s
3D Printed Object 19s 20s 29s 8s
Mustard 33s 9s 18s 19s
Garbage Bag Box 8s 8s 30s 7s
Average 24.9±3.624.9\pm 3.6s 14.8±1.714.8\pm 1.7s 22.1±2.322.1\pm 2.3s 16.1±1.816.1\pm 1.8s

Appendix I Finetuning

Finetuning is evaluated as a complementary final stage, with explicit learning-rate reduction and checkpoint selection.

I.1 Implementation

Finetuning reduces learning rate ten- to twentyfold and saves checkpoints frequently because excessive updates can hurt. Simulation evaluates every checkpoint and reports the best; hardware screens checkpoints with five to ten trials before fully evaluating the selected one. Results therefore include selection effects.

I.2 Table Cleaning (OXE)

For fifty cleaning demonstrations plus custom OXE, Figure 23 shows approximately .71 for co-training before and after finetuning, and .83 to .82 for Ambient with locality. The paper reports no significant gain; this is not evidence of exact equivalence under every finetuning schedule.

Complete original figure with every panel, label and legend.
Figure 23: Finetuning resulted in no statistically significant change in policy performance on the table cleaning task.

Appendix J Distribution Shifts in the Observation Space

Action-distribution alignment does not directly address mismatched observations. The appendix explores two approaches without finding improvements in the tested cleaning setting; broader observation-shift handling remains open.

J.1 Noising the Observations

Half of training examples receive observation noise on the same schedule as action noise, with both noise levels supplied to the model. Inference uses clean observations. This did not improve table-cleaning performance; it differs in purpose from concurrent observation-noising work aimed at exploration for RL finetuning.

J.2 Classifier-Free Guidance

Classifier-free guidance trains conditional and observation-free denoisers, then extrapolates between their outputs with a guidance weight. Equation 87 retains the exact rule. Sweeping that weight did not improve the tested cleaning policy, so the presence of an unconditional branch did not solve the observed distribution shift.

Equation 87
h~θ​(At,O,t)=(1+w)⋅hθ​(At,O,t)−w⋅hθ​(At,∅,t)\tilde{h}_{\theta}(A_{t},O,t)=(1+w)\cdot h_{\theta}(A_{t},O,t)-w\cdot h_{\theta}(A_{t},\varnothing,t) (87)
Complete original figure with every panel, label and legend.
Figure 24: Algorithms for handling distribution shift in the observation space. Neither approach produced stronger policies on the table cleaning task. (24(a)) Adding Gaussian noise to the observations during training. (24(b)) Classifier-free guidance [94] swept over guidance weights ww.
Selectable original panel labels

(a) Noising the observations during training.

(b) Classifier-free guidance vs. guidance weight ww.

Concept demonstration

Open image ↗