Literature
Display
← All papers

ZETA: A Controlled Study of Zero-Shot Cross-Embodiment VLA Transfer for Tabletop Manipulation

Mi Yan · Wenhao Zhang · Zhiqi Zhang · Yu Peng · Tangxinyu Wang · Lingfei Zhai · Jiayi Su · Shengliang Deng · Lin Peng · Yaowei Liu · Yuxing Chen · Zhiyuan Wei · Jilong Wang · Jiayi Chen · Jiangran Lyu · Zhizheng Zhang · He Wang

Preprint posted · CoRL 2026 (accepted) ↗

Robotics11 min text · CondensedOriginal paper ↗
Contents

Overview

This edition follows the 28-page arXiv v3 dated 11 October 2026, retaining all nine figures, 25 numbered tables, 15 displayed equations and 73 references. Four figures are restored from native PDF artwork to preserve fine labels, plots and original image detail. Read the scores as task progress, including half-credit contact, rather than binary success rates. The study concerns stationary tabletop arms with two-finger grippers; mobile platforms do not expose base control to the policy.

Original S0.F1 native PDF artwork with complete original panels, labels and layout.
Figure 1: Overview of our controlled study. We distinguish strict zero-shot transfer, where the target robot is absent from all training data, from pretrain-exposed zero-shot transfer. Under a controlled benchmark, we conduct fine-grained analyses across four embodiment shift categories, covering four research questions (RQ) on state-action representations, source-embodiment diversity, auxiliary co-training, and target-embodiment exposure.

At a glance

ZETA separates strict transfer to a robot absent from all training from transfer to a robot seen during pretraining but absent from downstream demonstrations. Across seven simulated and seven physical targets, controlled experiments favor local end-effector representations, broad procedural source diversity and auxiliary supervision. In simulation, 5% target-robot pretraining exposure adds 13.4 percentage points of average progress on two targets. These findings concern the tested tabletop setting, not universal transfer across robot morphologies.

1 Introduction

Comparisons of cross-embodiment policies can mix changes in tasks, cameras, objects, data quantity and target exposure. ZETA holds these factors controlled where possible and studies appearance, gripper, arm and combined shifts. Its four questions concern representations, source diversity, auxiliary objectives and target exposure. The first three use strict zero-shot protocols; the fourth deliberately includes the target during pretraining. The primary evidence is simulated, with physical validation of representations and diversity.

2 Related Work

Prior work supplies broad robot datasets, transferable policy interfaces and cross-embodiment evaluations. ZETA’s contribution is controlled comparison and explicit exposure accounting within a fixed manipulation regime.

2.1 Data Composition in Multi-Embodiment Pretraining

Pooling heterogeneous robots can improve coverage, but source diversity is often entangled with dataset scale, tasks and architecture. ZETA compares procedural source counts under both fixed-total and fixed-per-source trajectory budgets, then separately replaces source data with target data at fixed total size. Its Franka-family pool is narrower than an arbitrary collection of unrelated robots.

2.2 Cross-Embodiment Policy Design: Representations, Interfaces, and Supervision

Shared action spaces, embodiment-specific interfaces and auxiliary semantic supervision offer distinct transfer mechanisms. ZETA compares Cartesian state/action frames and auxiliary objectives under a shared backbone. Its language-action variant follows LAP; the state-free interpretation connects to work questioning explicit proprioceptive state inputs.

2.3 Zero-Shot Cross-Embodiment Evaluation and Benchmarks

An unseen downstream robot–task pair is different from a robot absent from all training. The benchmark separates those conditions and controls tasks, scenes, cameras and budgets, rather than ranking published methods whose exposure protocols differ.

3 Problem Formulation and Preliminaries

Pretraining contains a set of robots and tasks; post-training contains downstream tasks demonstrated only by a source robot. Evaluation uses a different target robot on those same downstream tasks. Strict zero-shot excludes that target from pretraining as well. Pretrain-exposed zero-shot permits target data only on pretraining tasks, which remain disjoint from downstream tasks. Appearance, gripper, arm and full changes identify which robot factors vary.

4 Study Design

The experiment varies robot representation, procedural training coverage, auxiliary labels or target exposure while retaining a common two-finger tabletop control interface.

4.1 Representative State-Action Representations

The four variants combine base-frame or current-gripper-frame motion deltas with absolute base-frame pose history or EEF-relative pose history. Joint-indexed interfaces are excluded because dimensions, limits and semantics differ across robots. The transform is vectorized as translation plus Euler angles. Expressing motion and history locally reduces dependence on a robot’s base coordinates; a zero-length relative-state history degenerates to a constant identity and carries no pose information.

Table 1: State and action representations studied in RQ1.
Action representation
World-Delta action {𝒱⁡(Tb,et′+1)−𝒱⁡(Tb,et′)}t′=tt+K−1\{\mathcal{V}(T_{b,e_{t^{\prime}+1}})-\mathcal{V}(T_{b,e_{t^{\prime}}})\}_{t^{\prime}=t}^{t+K-1}
EEF-Delta action {𝒱(Tb,et′−1⋅Tb,et′+1)=𝒱(Tet′,et′+1)}t′=tt+K−1\{\mathcal{V}(T_{b,e_{t^{\prime}}}^{-1}\cdot T_{b,e_{t^{\prime}+1}})=\mathcal{V}(T_{e_{t^{\prime}},e_{t^{\prime}+1}})\}_{t^{\prime}=t}^{t+K-1}
State representation
Abs. EEF state {𝒱⁡(Tb,et−i)}i=0h\{\mathcal{V}(T_{b,e_{t-i}})\}_{i=0}^{h}
EEF-Delta state {𝒱(Tb,et−1⋅Tb,et−i)=𝒱(Tet,et−i)}i=0h\{\mathcal{V}(T_{b,e_{t}}^{-1}\cdot T_{b,e_{t-i}})=\mathcal{V}(T_{e_{t},e_{t-i}})\}_{i=0}^{h}

4.2 Controlled Data Construction

Most comparisons pretrain on 640K pick-and-place trajectories, then use 40K source demonstrations per simulated downstream task or 50 per physical task. The 512 sources are procedurally varied Franka-style arms; commercial test robots are excluded under the strict protocol. Pretraining randomizes geometry, appearance, cameras, objects and simple physical parameters. The fixed-per-source diversity sweep instead allows the total pretraining dataset to grow. Physical validation uses real source post-training data, so it is not a test of direct simulation-only deployment.

Refer to caption
Figure 2: Procedurally generated embodiment pool for pretraining.

4.3 Auxiliary Co-training Objectives

Three auxiliary tasks predict structured language actions, the next manipulation-relevant goal pose or instruction-relevant object boxes. They supply local motion, spatial progress or grounding supervision through text tokens. Auxiliary training is enabled only in designated variants and is applied during both pretraining and post-training.

Original S4.F3 native PDF artwork with complete original panels, labels and layout.
Figure 3: Evaluation setup for controlled cross-embodiment transfer. We evaluate seven held-out target embodiments across four shift categories in simulation and the real world, carefully matching post-training and test scenes to align non-embodiment factors (cameras, environments, and tasks).

4.4 Embodiment Shift Categories

Simulation uses a Franka arm/hand source and seven held-out targets: two texture changes, one gripper replacement, two arm replacements and two combined changes. Physical experiments use Franka–Robotiq as the source and corresponding hardware categories. Keeping a two-finger interface makes comparison tractable but excludes dexterous-hand transfer.

5 Controlled Experiments

Each research question changes one designated factor or budget regime. Results are grouped by embodiment-shift category; the controlled training and evaluation setup is essential to interpreting their differences.

5.1 Shared Experimental Setup

The backbone follows π0.5. Evaluation matches post-training objects, scene, lighting and a 30×40 cm initial-pose region; wrist and third-person views are aligned across robots. Each simulated model has three independent training runs and 100 rollouts for each of three tasks on seven targets, totaling 6,300. Physical models have ten trials for each of two tasks on seven targets, totaling 140. Scores are macro-averaged task progress; simulation uncertainty is standard deviation across three training runs.

Table 2: RQ1 zero-shot simulation results: EEF-Delta states improve arm-only and full-embodiment transfer, while EEF-Delta actions yield better average score.
State Action 𝐞𝐬\mathbf{e_{s}} APP GRP ARM FULL Average
Abs. EEF World-Delta 87.7 ±\pm 2.1 86.3 ±\pm 1.3 82.6 ±\pm 1.7 38.5 ±\pm 0.8 33.9 ±\pm 2.3 60.3 ±\pm 1.1
Abs. EEF EEF-Delta 92.6 ±\pm 1.0 91.8 ±\pm 1.1 88.2 ±\pm 2.6 39.9 ±\pm 0.5 38.6 ±\pm 0.8 64.6 ±\pm 0.4
EEF-Delta World-Delta 87.8 ±\pm 1.4 88.0 ±\pm 0.9 77.2 ±\pm 1.0 69.8 ±\pm 1.9 58.4 ±\pm 1.3 73.4 ±\pm 0.7
EEF-Delta EEF-Delta 91.5 ±\pm 0.6 90.0 ±\pm 1.1 78.3 ±\pm 2.4 74.3 ±\pm 0.9 60.4 ±\pm 0.9 75.7 ±\pm 1.3

5.2 RQ1: How Do State-Action Representations Affect Transfer?

With all 512 sources, EEF-relative state is the major improvement under arm and full shifts: arm progress rises from roughly 39% to 70–74%, and full-shift progress from 34–39% to 58–60%. Combining EEF-relative state and action gives 75.7% average simulated progress versus 60.3% for absolute state/base-frame deltas. The same combination reaches 89.9% physical progress versus 56.0–61.6% for the alternatives. This best-average representation is used in the later diversity and auxiliary studies.

Table 3: Real-world results for RQ1. Each entry averages task progress on open the fryer and water the flower.
State Action 𝐞𝐬\mathbf{e_{s}} APP GRP ARM FULL Average
Abs. EEF World-Delta 87.5 87.5 67.5 47.5 21.3 56.0
Abs. EEF EEF-Delta 90.0 90.0 80.0 60.0 16.3 61.6
EEF-Delta World-Delta 85.0 77.5 60.0 62.5 43.1 60.8
EEF-Delta EEF-Delta 100.0 97.5 92.5 87.5 81.9 89.9
Original S5.F4 native PDF artwork with complete original panels, labels and layout.
Figure 4: RQ2: effect of source-embodiment diversity under a fixed pretraining budget. The 512-source setting improves over the single-source setting in both simulation (a) and the real world (b). Embodiment-diverse pretraining also shifts attention from diffuse robot–scene regions toward task-relevant objects and the gripper.
Refer to caption

(a) Simulation
Refer to caption
(b) Real world

Figure 5: RQ2 under a fixed per-embodiment pretraining budget. We sweep the number of source embodiments among 32, 128, and 512 while keeping the per-embodiment trajectory count fixed, so the total pretraining budget grows with |ℰpre||\mathcal{E}_{\mathrm{pre}}|. Average zero-shot progress is reported in simulation (a) and real-world evaluation (b).

5.3 RQ2: How Does Source-Embodiment Diversity Affect Transfer Under Different Budget Controls?

At fixed 640K trajectories, growing from one source to 512 reduces data per source from 640K to 1,250 but improves average simulated progress by about 18 points. Physical performance is non-monotonic: eight sources hurt arm/full transfer before 512 recover. A separate 32/128/512-source sweep fixes trajectories per source, so both coverage and total data grow; progress improves over that range. These are complementary budget comparisons, not a universal monotonic scaling law. Attention maps and feature overlap are qualitative supporting evidence.

5.4 RQ3: How Do Co-training Tasks Affect Transfer?

Auxiliary tasks are tested in simulation, where the required extra annotations are available. Mean progress rises from 75.7% without co-training to 77.7–82.3%; object-box prediction is strongest in the aggregate. EEF-only language-action labels outperform mixed-frame labels, consistent with the local action interface. Gains are concentrated in larger robot shifts, but these auxiliary effects have no matching physical trial set here.

Table 4: RQ3 simulation results: auxiliary co-training objectives. All co-training variants improve over imitation-only training, with the largest gains under larger embodiment shifts.
Post train data ese_{s} APP GRP ARM FULL Average
No co-training 91.5 ±\pm 0.6 90.0 ±\pm 1.1 78.3 ±\pm 2.4 74.3 ±\pm 0.9 60.4 ±\pm 0.9 75.7 ±\pm 1.3
Language-action (eef-frame) 90.5 ±\pm 1.1 88.6 ±\pm 0.2 86.7 ±\pm 1.4 79.3 ±\pm 0.9 70.5 ±\pm 0.5 81.3 ±\pm 0.5
Language-action (both frames) 89.5 ±\pm 1.5 88.6 ±\pm 0.4 79.8 ±\pm 0.8 73.8 ±\pm 1.6 68.4 ±\pm 1.5 77.7 ±\pm 0.3
Subgoal 89.1 ±\pm 1.8 90.3 ±\pm 0.8 87.4 ±\pm 2.2 77.8 ±\pm 0.5 71.5 ±\pm 1.5 81.8 ±\pm 0.8
Task-conditioned BBox 90.4 ±\pm 1.4 90.2 ±\pm 0.4 87.0 ±\pm 0.9 80.1 ±\pm 0.8 71.9 ±\pm 1.4 82.3 ±\pm 0.2

5.5 RQ4: How Does Target-Embodiment Exposure Affect Transfer?

For UR5e–UMI and GoogleRobot, target trajectories replace source trajectories within a fixed 640K pretraining budget. Five-percent exposure raises the reported per-target means from 69.3% to 78.6% and 51.4% to 68.9%, averaging a 13.4-point gain. Target downstream demonstrations remain excluded; an additional target-post-trained oracle is explicitly non-zero-shot. Table 5 reports an interaction with representation and narrower gaps after exposure. Its EEF/EEF value at 5% exposure is 71.1%, below the 73.75% average of the two Figure 6 means; both original results are retained without assuming an unstated reconciliation.

Refer to caption

(a) RQ4 pretraining composition setting

Refer to caption

(b) UR5eUMI

Refer to caption

(c) Google Robot

Figure 6: RQ4: target-embodiment exposure during pretraining. Only 5% of target embodiment pretraining data substantially reduce the transfer gap on both UR5eUMI and Google Robot, but remain below the oracle performance of post-training on the target embodiment and test tasks.
Table 5: Interaction between state-action representation and target-embodiment exposure. Average progress (%) on the RQ4 targets.
State Action 0% 5%
Abs. EEF World-Delta 33.9 66.9
Abs. EEF EEF-Delta 38.6 69.2
EEF-Delta World-Delta 58.4 67.5
EEF-Delta EEF-Delta 60.4 71.1

5.6 Practical Guidance for Cross-Embodiment Transfer within the Studied Setting

Within this setting, report target exposure explicitly, favor local EEF state/actions for large robot shifts, evaluate source diversity together with its data allocation, and consider auxiliary supervision. Absolute state can remain competitive for smaller appearance or gripper changes. The recommendations are conditional on the tested backbone, tasks and procedural source family.

6 Limitations and Conclusion

The controlled setting isolates robot differences but does not test mobile-base control, dexterous hands or long-horizon tasks. Matched objects and camera layouts deliberately remove additional generalization challenges. Source diversity, representations and auxiliary labels improve transfer here; target exposure changes the problem enough that strict and pretrain-exposed results should be reported separately. Deployment engineering and broad morphology coverage remain open.

Appendix A Formal Definitions of the Two Zero-Shot Protocols

Both protocols post-train only on the source robot’s downstream demonstrations. Strict zero-shot excludes the target robot everywhere; pretrain-exposed zero-shot includes that robot on disjoint pretraining tasks. Neither uses target downstream demonstrations, but their hardware familiarity and difficulty differ. The retained formal table makes this distinction explicit.

Table 6: Formal distinction between the two zero-shot cross-embodiment transfer protocols.
Protocol Pre-train data Post-train data Eval.
Strict zero-shot
𝒟pre=𝒟⁡(ℰpre,𝒯pre)\mathcal{D}_{\mathrm{pre}}=\mathcal{D}(\mathcal{E}_{\mathrm{pre}},\mathcal{T}_{\mathrm{pre}})
𝒆𝒕∉𝓔𝐩𝐫𝐞e_{t}\notin\mathcal{E}_{\mathrm{pre}}
𝒯pre∩𝒯post=∅\mathcal{T}_{\mathrm{pre}}\cap\mathcal{T}_{\mathrm{post}}=\varnothing
𝒟post=𝒟⁡({es},𝒯post)\mathcal{D}_{\mathrm{post}}=\mathcal{D}(\{e_{s}\},\mathcal{T}_{\mathrm{post}})
es≠ete_{s}\neq e_{t}
emb.: ete_{t}
task: 𝒯post\mathcal{T}_{\mathrm{post}}
Pretrain- exposed zero-shot
𝒟pre=𝒟⁡(ℰpre,𝒯pre)\mathcal{D}_{\mathrm{pre}}=\mathcal{D}(\mathcal{E}_{\mathrm{pre}},\mathcal{T}_{\mathrm{pre}})
𝒆𝒕∈𝓔𝐩𝐫𝐞e_{t}\in\mathcal{E}_{\mathrm{pre}}
𝒯pre∩𝒯post=∅\mathcal{T}_{\mathrm{pre}}\cap\mathcal{T}_{\mathrm{post}}=\varnothing
𝒟post=𝒟⁡({es},𝒯post)\mathcal{D}_{\mathrm{post}}=\mathcal{D}(\{e_{s}\},\mathcal{T}_{\mathrm{post}})
es≠ete_{s}\neq e_{t}
emb.: ete_{t}
task: 𝒯post\mathcal{T}_{\mathrm{post}}

Appendix B Per-RQ Experimental Design Matrix

The design matrix records each question’s variable, exposure protocol, data pool, representation and evaluation platform. Representation and diversity have simulated and physical evidence; auxiliary objectives and target-exposure comparisons are simulated. Shared backbone, observations, action horizon, post-training data and scenes are fixed unless the relevant question explicitly varies them.

Table 7: Per-RQ experimental design matrix. The table summarizes the protocol, pretraining data, representation choice, and evaluation platform used to isolate each experimental factor.
RQ Factor studied Zero-shot protocol Pretraining data Representation Eval.
RQ1 State-action representation Strict zero-shot 512 non-target source embodiments Varied across four designs Sim. + real
RQ2 Source-embodiment diversity Strict zero-shot Primary: 1/8/512 sources, fixed total budget. Complementary: 32/128/512 sources, fixed per-embodiment budget. Fixed to best-average RQ1 Sim. + real
RQ3 Auxiliary co-training objective Strict zero-shot 512 non-target source embodiments with auxiliary objectives Fixed to best-average RQ1 Sim.
RQ4 Target-embodiment exposure Pretrain-exposed zero-shot Fixed-size pretraining pool with controlled target-embodiment replacement Fixed to best-average RQ1 Sim.

Appendix C Detailed Experimental Results

Tables 8–19 retain simulated per-task/per-robot results, including the source robot and all seven held-out targets; entries list the three training runs separately. Tables 20–23 retain physical representation/diversity results for opening the fryer and watering the flower, with ten trials per cell. These detailed results expose task-specific failures and non-monotonic diversity effects hidden by aggregate category means.

Table 8: Detailed RQ1 simulation task progress for serving sausages. Rows report individual robots, while columns vary the state/action representation. Each entry reports run-1/ run-2/ run-3 task progress (%).
Target embodiment Abs. EEF / World-Delta Abs. EEF / EEF-Delta EEF-Delta / World-Delta EEF-Delta / EEF-Delta
Franka 80.8/ 83.8/ 79.8 89.9/ 82.8/ 89.9 85.9/ 80.8/ 86.9 84.8/ 91.9/ 90.9
Franka+logo 83.8/ 79.8/ 82.8 83.8/ 90.9/ 93.9 86.9/ 85.9/ 86.9 88.9/ 85.9/ 86.9
Franka+green 84.8/ 83.8/ 82.8 87.9/ 88.9/ 89.9 81.8/ 82.8/ 86.9 85.9/ 91.9/ 86.9
Franka+UMI 79.8/ 78.8/ 77.8 83.8/ 90.9/ 83.8 70.7/ 80.8/ 76.8 74.8/ 81.8/ 86.9
UR5e+Franka 65.7/ 68.7/ 70.7 70.7/ 69.7/ 70.7 76.8/ 81.8/ 74.8 78.8/ 81.8/ 75.8
Google+Franka 0.0/ 0.0/ 0.0 2.0/ 1.0/ 1.0 54.5/ 53.5/ 53.5 50.5/ 58.6/ 56.6
UR5e+UMI 69.7/ 60.6/ 58.6 69.7/ 76.8/ 75.8 69.7/ 62.7/ 59.6 69.7/ 69.7/ 69.7
GoogleRobot 1.0/ 1.0/ 0.0 0.0/ 1.0/ 0.0 59.6/ 58.6/ 64.7 54.5/ 55.6/ 53.5
Table 9: Detailed RQ1 simulation task progress for stacking bowls. Rows report individual robots, while columns vary the state/action representation. Each entry reports run-1/ run-2/ run-3 task progress (%).
Target embodiment Abs. EEF / World-Delta Abs. EEF / EEF-Delta EEF-Delta / World-Delta EEF-Delta / EEF-Delta
Franka 82.0/ 90.0/ 87.0 95.0/ 93.0/ 89.0 84.0/ 84.0/ 88.0 91.0/ 87.0/ 86.0
Franka+logo 87.0/ 83.0/ 83.0 95.0/ 87.0/ 89.0 87.0/ 86.0/ 89.0 79.0/ 88.0/ 87.0
Franka+green 87.0/ 89.0/ 78.0 95.0/ 85.0/ 85.0 82.0/ 82.0/ 84.0 87.0/ 89.0/ 90.0
Franka+UMI 77.0/ 83.0/ 77.0 86.0/ 86.0/ 81.0 82.0/ 73.0/ 75.0 73.0/ 77.0/ 69.0
UR5e+Franka 78.0/ 75.0/ 77.0 71.0/ 79.0/ 73.0 72.0/ 68.0/ 65.0 83.0/ 83.0/ 85.0
Google+Franka 0.0/ 0.0/ 2.0 0.0/ 0.0/ 0.0 64.0/ 66.0/ 58.0 77.0/ 70.0/ 73.0
UR5e+UMI 70.0/ 64.0/ 61.0 77.0/ 65.0/ 71.0 62.0/ 53.0/ 50.0 74.0/ 74.0/ 69.0
GoogleRobot 0.0/ 1.0/ 0.0 0.0/ 0.0/ 0.0 41.0/ 55.0/ 48.0 29.0/ 54.0/ 54.0
Table 10: Detailed RQ1 simulation task progress for stacking cubes. Rows report individual robots, while columns vary the state/action representation. Each entry reports run-1/ run-2/ run-3 task progress (%).
Target embodiment Abs. EEF / World-Delta Abs. EEF / EEF-Delta EEF-Delta / World-Delta EEF-Delta / EEF-Delta
Franka 94.0/ 98.0/ 94.0 97.0/ 99.0/ 98.0 97.0/ 93.0/ 91.0 98.0/ 98.0/ 96.0
Franka+logo 92.0/ 97.0/ 90.0 99.0/ 98.0/ 98.0 95.0/ 93.0/ 93.0 96.0/ 98.0/ 96.0
Franka+green 87.0/ 92.0/ 90.0 96.0/ 92.0/ 98.0 97.0/ 91.0/ 93.0 94.0/ 93.0/ 96.0
Franka+UMI 95.0/ 89.0/ 86.0 93.0/ 98.0/ 91.0 82.0/ 79.0/ 75.7 77.0/ 81.0/ 84.0
UR5e+Franka 79.0/ 82.0/ 81.0 89.0/ 92.0/ 91.0 75.0/ 88.0/ 79.0 71.0/ 77.0/ 84.0
Google+Franka 2.0/ 10.0/ 2.0 3.0/ 2.0/ 3.0 79.0/ 74.0/ 74.0 79.0/ 75.0/ 78.0
UR5e+UMI 79.0/ 71.0/ 67.0 84.0/ 80.0/ 85.0 59.0/ 58.0/ 61.0 69.0/ 63.0/ 66.0
GoogleRobot 2.0/ 2.0/ 2.0 6.0/ 2.0/ 2.0 70.0/ 57.0/ 63.0 59.0/ 47.0/ 56.0
Table 11: Detailed RQ2 simulation task progress for serving sausages. Rows report individual robots, while columns vary the number of non-target source embodiments in the fixed-budget pretraining pool. Each entry reports run-1/ run-2/ run-3 task progress (%).
Target embodiment 1 source 8 sources 512 sources
Franka 84.8/ 80.8/ 81.8 82.8/ 81.8/ 85.9 84.8/ 91.9/ 90.9
Franka+logo 80.8/ 76.8/ 84.8 84.8/ 80.8/ 81.8 88.9/ 85.9/ 86.9
Franka+green 78.8/ 81.8/ 79.8 80.8/ 84.8/ 85.9 85.9/ 91.9/ 86.9
Franka+UMI 53.5/ 53.5/ 61.6 72.7/ 75.8/ 77.8 74.8/ 81.8/ 86.9
UR5e+Franka 40.4/ 50.5/ 45.5 52.5/ 52.5/ 58.6 78.8/ 81.8/ 75.8
Google+Franka 41.4/ 45.5/ 45.5 56.6/ 41.4/ 53.5 50.5/ 58.6/ 56.6
UR5e+UMI 40.4/ 33.3/ 40.4 63.6/ 54.5/ 56.6 69.7/ 69.7/ 69.7
GoogleRobot 31.3/ 28.3/ 26.3 47.5/ 45.5/ 43.4 54.5/ 55.6/ 53.5
Table 12: Detailed RQ2 simulation task progress for stacking bowls. Rows report individual robots, while columns vary the number of non-target source embodiments in the fixed-budget pretraining pool. Each entry reports run-1/ run-2/ run-3 task progress (%).
Target embodiment 1 source 8 sources 512 sources
Franka 87.0/ 81.0/ 84.0 92.0/ 87.0/ 85.0 91.0/ 87.0/ 86.0
Franka+logo 84.0/ 84.0/ 83.0 89.0/ 87.0/ 93.0 79.0/ 88.0/ 87.0
Franka+green 85.0/ 85.0/ 86.0 86.0/ 86.0/ 86.0 87.0/ 89.0/ 90.0
Franka+UMI 64.0/ 65.0/ 58.0 68.0/ 71.0/ 67.0 73.0/ 77.0/ 69.0
UR5e+Franka 57.0/ 60.0/ 56.0 76.0/ 73.0/ 76.0 83.0/ 83.0/ 85.0
Google+Franka 55.0/ 64.0/ 61.0 70.0/ 71.0/ 62.0 77.0/ 70.0/ 73.0
UR5e+UMI 51.0/ 49.0/ 39.0 63.0/ 66.0/ 64.0 74.0/ 74.0/ 69.0
GoogleRobot 15.0/ 28.0/ 29.0 45.0/ 35.0/ 49.0 29.0/ 54.0/ 54.0
Table 13: Detailed RQ2 simulation task progress for stacking cubes. Rows report individual robots, while columns vary the number of non-target source embodiments in the fixed-budget pretraining pool. Each entry reports run-1/ run-2/ run-3 task progress (%).
Target embodiment 1 source 8 sources 512 sources
Franka 93.0/ 92.0/ 91.0 95.0/ 95.0/ 96.0 98.0/ 98.0/ 96.0
Franka+logo 89.0/ 87.0/ 93.0 94.0/ 98.0/ 95.0 96.0/ 98.0/ 96.0
Franka+green 91.0/ 89.0/ 87.0 96.0/ 97.0/ 98.0 94.0/ 93.0/ 96.0
Franka+UMI 53.0/ 51.0/ 59.0 86.0/ 81.0/ 87.0 77.0/ 81.0/ 84.0
UR5e+Franka 55.0/ 59.0/ 60.0 68.0/ 69.0/ 63.0 71.0/ 77.0/ 84.0
Google+Franka 60.0/ 64.0/ 59.0 66.0/ 69.0/ 67.0 79.0/ 75.0/ 78.0
UR5e+UMI 39.0/ 36.0/ 33.0 61.0/ 55.0/ 55.0 69.0/ 63.0/ 66.0
GoogleRobot 39.0/ 35.0/ 28.0 43.0/ 46.0/ 52.0 59.0/ 47.0/ 56.0
Table 14: Detailed RQ3 simulation task progress for serving sausages. Rows report individual robots, while columns vary the auxiliary co-training objective. Each entry reports run-1/ run-2/ run-3 task progress (%).
Target embodiment No co-training LAP (EEF frame) LAP (both frames) Subgoal BBox
Franka 84.8/ 91.9/ 90.9 90.9/ 84.8/ 91.9 88.9/ 91.9/ 91.9 91.9/ 95.0/ 93.9 85.9/ 91.9/ 87.9
Franka+logo 88.9/ 85.9/ 86.9 87.9/ 85.9/ 86.9 85.9/ 87.9/ 90.9 97.0/ 91.9/ 95.0 92.9/ 92.9/ 91.9
Franka+green 85.9/ 91.9/ 86.9 80.8/ 87.9/ 80.8 87.9/ 86.9/ 89.9 92.9/ 96.0/ 95.0 88.9/ 88.9/ 88.9
Franka+UMI 74.8/ 81.8/ 86.9 84.8/ 82.8/ 85.9 76.8/ 79.8/ 76.8 84.8/ 88.9/ 93.9 85.9/ 86.9/ 88.9
UR5e+Franka 78.8/ 81.8/ 75.8 78.8/ 84.8/ 81.8 72.7/ 76.8/ 68.7 89.9/ 91.9/ 92.9 84.8/ 83.8/ 83.8
Google+Franka 50.5/ 58.6/ 56.6 77.8/ 74.8/ 73.7 58.6/ 58.6/ 61.6 68.7/ 69.7/ 70.7 71.7/ 72.7/ 64.7
UR5e+UMI 69.7/ 69.7/ 69.7 77.8/ 72.7/ 74.8 70.7/ 74.8/ 70.7 80.8/ 78.8/ 79.8 79.8/ 85.9/ 73.7
GoogleRobot 54.5/ 55.6/ 53.5 67.7/ 61.6/ 68.7 62.6/ 50.5/ 55.6 62.6/ 62.6/ 64.7 59.6/ 69.7/ 57.6
Table 15: Detailed RQ3 simulation task progress for stacking bowls. Rows report individual robots, while columns vary the auxiliary co-training objective. Each entry reports run-1/ run-2/ run-3 task progress (%).
Target embodiment No co-training LAP (EEF frame) LAP (both frames) Subgoal BBox
Franka 91.0/ 87.0/ 86.0 84.0/ 87.0/ 89.0 81.0/ 86.0/ 87.0 80.0/ 82.0/ 86.0 87.0/ 90.0/ 86.0
Franka+logo 79.0/ 88.0/ 87.0 85.0/ 89.0/ 84.0 92.0/ 89.0/ 84.0 85.0/ 79.0/ 84.0 81.0/ 89.0/ 85.0
Franka+green 87.0/ 89.0/ 90.0 88.0/ 85.0/ 85.0 82.0/ 82.0/ 83.0 85.0/ 76.0/ 84.0 89.0/ 83.0/ 86.0
Franka+UMI 73.0/ 77.0/ 69.0 86.0/ 80.0/ 82.0 74.0/ 74.0/ 75.0 78.0/ 82.0/ 86.0 79.0/ 79.0/ 82.0
UR5e+Franka 83.0/ 83.0/ 85.0 83.0/ 86.0/ 80.0 73.0/ 84.0/ 82.0 83.0/ 80.0/ 86.0 79.0/ 76.0/ 79.0
Google+Franka 77.0/ 70.0/ 73.0 69.0/ 68.0/ 69.0 76.0/ 72.0/ 61.0 73.0/ 68.0/ 60.0 76.0/ 83.0/ 74.0
UR5e+UMI 74.0/ 74.0/ 69.0 73.0/ 80.0/ 71.0 70.0/ 65.0/ 68.0 83.0/ 72.0/ 81.0 74.0/ 72.0/ 80.0
GoogleRobot 29.0/ 54.0/ 54.0 65.0/ 68.0/ 58.0 62.0/ 55.0/ 65.0 31.0/ 55.0/ 57.0 64.0/ 67.0/ 64.0
Table 16: Detailed RQ3 simulation task progress for stacking cubes. Rows report individual robots, while columns vary the auxiliary co-training objective. Each entry reports run-1/ run-2/ run-3 task progress (%).
Target embodiment No co-training LAP (EEF frame) LAP (both frames) Subgoal BBox
Franka 98.0/ 98.0/ 96.0 96.0/ 96.0/ 95.0 93.0/ 91.0/ 95.0 89.0/ 90.0/ 94.0 96.0/ 95.0/ 94.0
Franka+logo 96.0/ 98.0/ 96.0 94.0/ 94.0/ 96.0 91.0/ 94.0/ 95.0 94.0/ 94.0/ 93.0 96.0/ 95.0/ 94.0
Franka+green 94.0/ 93.0/ 96.0 97.0/ 91.0/ 97.0 92.0/ 89.0/ 92.0 92.0/ 98.0/ 94.0 96.0/ 94.0/ 92.0
Franka+UMI 77.0/ 81.0/ 84.0 95.0/ 93.0/ 91.0 86.0/ 85.0/ 91.0 91.0/ 92.0/ 90.0 94.0/ 93.0/ 94.0
UR5e+Franka 71.0/ 77.0/ 84.0 92.0/ 90.0/ 88.0 86.0/ 88.0/ 83.0 80.0/ 83.0/ 83.0 91.0/ 84.0/ 89.0
Google+Franka 79.0/ 75.0/ 78.0 74.0/ 79.0/ 77.0 68.0/ 77.0/ 82.0 73.0/ 77.0/ 70.0 84.0/ 80.0/ 85.0
UR5e+UMI 69.0/ 63.0/ 66.0 86.0/ 84.0/ 86.0 87.0/ 80.0/ 86.0 86.0/ 89.0/ 87.0 83.0/ 89.0/ 87.0
GoogleRobot 59.0/ 47.0/ 56.0 56.0/ 58.0/ 60.0 66.0/ 73.0/ 70.0 75.0/ 71.0/ 71.0 61.0/ 59.0/ 68.0
Table 17: Detailed RQ4 simulation task progress for serving sausages. Rows vary the target-embodiment pretraining exposure ratio and the target post-training oracle; columns report the two target embodiments used in the sweep. Each entry reports run-1/ run-2/ run-3 task progress (%).
Setting UR5e+UMI GoogleRobot
Strict zero-shot (0%) 69.7/ 69.7/ 69.7 54.5/ 55.6/ 53.5
Pretrain-exposed (5%) 75.8/ 73.7/ 74.8 68.7/ 64.7/ 68.7
Pretrain-exposed (30%) 86.9/ 83.8/ 75.8 71.7/ 72.7/ 67.7
Pretrain-exposed (100%) 89.9/ 84.8/ 85.9 73.7/ 68.7/ 71.7
Target post-training oracle 83.8/ 78.8/ 80.8 69.7/ 62.6/ 60.6
Table 18: Detailed RQ4 simulation task progress for stacking bowls. Rows vary the target-embodiment pretraining exposure ratio and the target post-training oracle; columns report the two target embodiments used in the sweep. Each entry reports run-1/ run-2/ run-3 task progress (%).
Setting UR5e+UMI GoogleRobot
Strict zero-shot (0%) 74.0/ 74.0/ 69.0 29.0/ 54.0/ 54.0
Pretrain-exposed (5%) 80.0/ 77.0/ 80.0 70.0/ 61.0/ 67.0
Pretrain-exposed (30%) 87.0/ 82.0/ 85.0 70.0/ 69.0/ 59.0
Pretrain-exposed (100%) 88.0/ 81.0/ 86.0 71.0/ 61.0/ 71.0
Target post-training oracle 97.0/ 92.0/ 92.0 81.0/ 79.0/ 76.0
Table 19: Detailed RQ4 simulation task progress for stacking cubes. Rows vary the target-embodiment pretraining exposure ratio and the target post-training oracle; columns report the two target embodiments used in the sweep. Each entry reports run-1/ run-2/ run-3 task progress (%).
Setting UR5e+UMI GoogleRobot
Strict zero-shot (0%) 69.0/ 63.0/ 66.0 59.0/ 47.0/ 56.0
Pretrain-exposed (5%) 87.0/ 79.0/ 80.0 75.0/ 68.0/ 77.0
Pretrain-exposed (30%) 91.0/ 90.0/ 87.0 76.0/ 68.0/ 67.0
Pretrain-exposed (100%) 92.0/ 84.0/ 84.0 77.0/ 71.0/ 68.0
Target post-training oracle 98.0/ 95.0/ 93.0 90.0/ 88.0/ 89.0
Table 20: Detailed RQ1 real-world task progress for open the fryer. Rows report individual robots, while columns vary the state/action representation. Each entry reports mean task progress (%) over 10 rollouts.
Target embodiment Abs. EEF / World-Delta Abs. EEF / EEF-Delta EEF-Delta / World-Delta EEF-Delta / EEF-Delta
Franka+Robotiq 90.0 95.0 80.0 100.0
Franka+Robotiq (CoRL logo) 90.0 90.0 90.0 100.0
Franka+Umi 80.0 85.0 90.0 100.0
UR 55.0 35.0 55.0 90.0
Humanoid robot 0.0 0.0 35.0 100.0
UR+Umi 60.0 50.0 60.0 80.0
Humanoid robot+Umi 0.0 0.0 55.0 85.0
Piper-on-Quadruped 35.0 0.0 40.0 80.0
Table 21: Detailed RQ1 real-world task progress for water the flower. Rows report individual robots, while columns vary the state/action representation. Each entry reports mean task progress (%) over 10 rollouts.
Target embodiment Abs. EEF / World-Delta Abs. EEF / EEF-Delta EEF-Delta / World-Delta EEF-Delta / EEF-Delta
Franka+Robotiq 85.0 85.0 90.0 100.0
Franka+Robotiq (CoRL logo) 85.0 90.0 65.0 95.0
Franka+Umi 55.0 75.0 30.0 85.0
UR 40.0 85.0 70.0 85.0
Humanoid robot 0.0 0.0 40.0 80.0
UR+Umi 55.0 60.0 60.0 75.0
Humanoid robot+Umi 0.0 0.0 25.0 65.0
Piper-on-Quadruped 20.0 20.0 30.0 90.0
Table 22: Detailed RQ2 real-world task progress for open the fryer. Rows report individual robots, while columns vary the number of pretraining embodiments |ℰpre||\mathcal{E}_{\mathrm{pre}}|. Each entry reports mean task progress (%) over 10 rollouts.
Target embodiment 1 source 8 sources 512 sources
Franka+Robotiq 80.0 100.0 100.0
Franka+Robotiq (CoRL logo) 90.0 100.0 100.0
Franka+Umi 80.0 100.0 100.0
UR 75.0 75.0 90.0
Humanoid robot 95.0 90.0 100.0
UR+Umi 65.0 75.0 80.0
Humanoid robot+Umi 75.0 70.0 85.0
Piper-on-Quadruped 90.0 45.0 80.0
Table 23: Detailed RQ2 real-world task progress for water the flower. Rows report individual robots, while columns vary the number of pretraining embodiments |ℰpre||\mathcal{E}_{\mathrm{pre}}|. Each entry reports mean task progress (%) over 10 rollouts.
Target embodiment 1 source 8 sources 512 sources
Franka+Robotiq 80.0 95.0 100.0
Franka+Robotiq (CoRL logo) 85.0 85.0 95.0
Franka+Umi 30.0 65.0 85.0
UR 75.0 35.0 85.0
Humanoid robot 85.0 30.0 80.0
UR+Umi 55.0 35.0 75.0
Humanoid robot+Umi 50.0 30.0 65.0
Piper-on-Quadruped 65.0 70.0 90.0

Appendix D Taxonomy of Zero-Shot Protocols in Prior Work

The source reclassifies representative prior papers by target exposure during pretraining and post-training. Some “zero-shot” claims concern only unseen robot–task combinations, while others exclude the target robot entirely. This is the authors’ protocol taxonomy, not a controlled rerun or performance ranking of those prior systems; original table qualifications and citations remain.

Table 24: Classification of representative prior cross-embodiment manipulation evaluations described as zero-shot or out-of-the-box in the literature, reclassified according to target-embodiment exposure under our protocol taxonomy.
Work Target embodiment Target embodiment Protocol
in pretraining? in post-training?
LAP [64] ✗ ✗ Strict
Cloak [39] ✗ ✗ Strict
RDT2 [29] ✗ ✗ Strict
Octo [15] ✓ ✗ Pretrain-exposed
π0.7\pi_{0.7} [37] ✓ ✗ Pretrain-exposed
Qwen-RobotManip [61] ✓ ✗ Pretrain-exposed
OpenVLA [20] ✓ ✗ Pretrain-exposed
Gemini Robotics 1.5 [14]∗ ✓ ✗ Pretrain-exposed
Being-H0.5 [31]† ✓ ✓ Unseen task–emb. pair

∗Gemini Robotics 1.5 reports zero-shot transfer to tasks observed only on another embodiment, while the evaluated target embodiments are already included in its multi-embodiment pretraining and receive no robot-specific post-training.

†Being-H0.5 reports unseen task–embodiment pairs as zero-shot even though the target embodiment participates in post-training, so the evaluation falls outside both strict and pretrain-exposed protocols.

Appendix E Cross-Embodiment Feature Visualization

Mean-pooled final VLM features from 3,600 frames in 1,200 trajectories are projected with cosine-distance UMAP. UR5e–UMI and GoogleRobot features overlap more visibly after 512-source than one-source training. This selected two-dimensional visualization is qualitative alignment evidence, not a quantitative causal explanation of transfer.

Refer to caption
Figure 7: UMAP of mean-pooled last-layer VLM features from held-out UR5eUMI and GoogleRobot observations.

Appendix F Data-Construction Details

Six movable arm-link regions are scaled roughly 0.7–1.3, with corresponding visual/collision geometry changes. Gripper fingers, contact planks and low-saturation 6–8-color palettes vary across 512 Franka-style sources. Held-out green/logo appearances have saturation 0.77/0.99 versus source maximum 0.46, but are not strongly out-of-distribution in all visual respects. Pretraining uses over 1,000 materials and 10,000 filtered objects, randomized lights/layouts/cameras, mass and friction. Post-training/evaluation deliberately keep each downstream scene and object set fixed.

Original A6.F8 native PDF artwork with complete original panels, labels and layout.
Figure 8: HSV color distribution of the procedurally generated source palette versus the appearance-only evaluation targets. The held-out green and logo-orange colors lie outside the source palette primarily along the saturation dimension.
Refer to caption
Figure 9: Empirical distributions of the procedurally generated source-embodiment pool. We summarize the arm-scale factors, sampled gripper dimensions, contact-plank geometry, and color-hue coverage for the 512 generated Franka-style tabletop embodiments used for source pretraining.

Appendix G Embodiment Shift Taxonomy

The following taxonomy identifies the exact arm, gripper and appearance changes. A mobile platform’s presence does not imply its base is controlled or represented by the policy.

G.1 Simulation

The simulated source is a seven-degree-of-freedom Franka arm with its standard hand. Targets add a logo or green fingers, attach a UMI gripper, pair a UR5e or GoogleRobot arm with a Franka hand, or change both arm and gripper. GoogleRobot’s mobile-base states and commands are excluded. The two appearance targets retain the source kinematics.

G.2 Real World

The physical source is Franka–Robotiq. Targets are logo-textured Franka–Robotiq, Franka–UMI, UR5e–Robotiq, UR5e–UMI, a humanoid with native or UMI gripper, and a Piper arm on a quadruped. Humanoid trials expose only the right arm and fix the left; mobile and quadruped base states/actions are excluded throughout. Thus “full” denotes combined manipulator/end-effector changes within the stationary task interface.

Appendix H Experimental Setup Details

The appendix specifies simulator/rendering choices, camera calibration, scene controls, progress scoring and downstream tasks needed to interpret the comparison.

H.1 Simulation Platform

MuJoCo supplies physics and IsaacSim supplies rendering for the simulated experiments.

H.2 Camera Calibration Details

A shared tabletop frame separates task geometry from robot base conventions. Third-person camera extrinsics are identical in that frame for corresponding simulated trajectories; wrist cameras are aligned relative to each gripper. Physical cameras approximate the same workspace and gripper coverage. These calibrations reduce viewpoint shift, so performance is not evidence for transfer to arbitrary uncalibrated views.

H.3 Environment and Object Set Details

Simulation fixes lighting, materials and workspace across corresponding robot trials. Physical experiments use a black enclosure and gray tablecloth with matched layout. Evaluation uses exactly the object instances from source post-training, excluding an additional object-identity shift.

H.4 Details of Metrics

Each rollout scores 0, 0.5 or 1: failure, contact with the necessary task object, or full completion. Means are expressed as percentages. Aggregation first averages tasks per robot, then robots within each of four shift categories, then the four categories equally. Because categories contain different robot counts, this is not a simple mean across all target robots. Simulation standard deviations summarize three run-level aggregates.

H.5 Downstream Tasks

Simulated downstream tasks serve sausages and stack bowls or cubes, with 40K source trajectories each. Physical tasks open an air fryer and water a plant, with 50 source demonstrations each. These downstream tasks are disjoint from pretraining tasks.

Appendix I Auxiliary Co-Training Objective Specifications

The loader samples imitation/auxiliary examples at 0.75/0.25. Actions use flow matching; auxiliary answers use next-token cross-entropy. Language actions convert translation to centimeters and rotation to degrees, round to integers and verbalize nonzero motion plus gripper state; mixed-frame prompts explicitly state their reference frame. Subgoals encode the next relevant waypoint in the current EEF frame. Bounding boxes are task-conditioned, one per camera, in XYXY order with normalized coordinates quantized into 256 bins. Exact prompts and response formats are retained in Table 25.

Table 25: Auxiliary co-training objective specifications. Braces denote fields filled from the trajectory annotation. The target column lists example answer formats.
Objective Prompt template Target format
Language-action (EEF frame) The task is {instruction}. Predict the robot’s action in the end-effector frame; State: {state tokens}; Answer: move {direction} {value} cm; tilt/rotate {direction} {value} degrees; open gripper; close gripper.
Language-action (mixed frame) The task is {instruction}. Predict the robot’s action in the {action frame}; State: {state tokens}; Answer: move {direction} {value} cm; rotate {direction} {value} degrees; close gripper.
Subgoal prediction In order to {instruction}, what is the next goal pose as xyz and rpy? Answer: [x,y,z,roll,pitch,yaw], e.g., [0.012,-0.004,0.085,1.57,0.00,-0.12].
Task-conditioned BBox In order to {instruction}, what is the bounding box of the object to be manipulated in the images? Answer: [x_min,y_min,x_max,y_max] per view, e.g., [[q_xmin,q_ymin,q_xmax,q_ymax], ...].

Appendix J Additional RQ4 Target-Exposure Details

Each exposure sweep builds a separate target-specific 640K pick-and-place pool and replaces a chosen fraction of non-target trajectories. The two targets differ in arm degrees of freedom and platform morphology, but base control remains excluded. The target-post-trained oracle uses downstream target demonstrations and serves only as an upper reference. Pretraining familiarity narrows, but does not remove, the remaining target-task adaptation gap.

Appendix K Model Training Details

A pretrained vision-language stream and separate action expert interact through a mixture-of-transformers backbone. Inputs use 224×224 images, four state-history entries and 99th-percentile normalization; outputs are 12-step action chunks. Flow matching trains actions, with an additional auxiliary cross-entropy weight of 0.25 when enabled. Pretraining runs 120K steps at batch 320 on 16 H800 GPUs; post-training runs 100K steps at batch 160 on eight. The separate loader ratio and auxiliary-loss weight are both reported source settings.

Appendix L Policy Inference and Action Execution Details

Each request supplies an instruction, wrist/third-person RGB images and four EEF states. The retained formulas give all four representation-specific request/response conversions: base-frame deltas update vectorized absolute pose, while local EEF deltas compose transforms. Gripper output is quantized to close/no-op/open. The server batches up to ten requests; one warmed-up query takes 168 ms and 6,168 MiB on an RTX 4090 in the reported setup. These are serving measurements, not a complete sensing-to-actuation latency or guaranteed control frequency.

Equations
A^t:t+11={a^j}j=tt+11,a^j=[δvj,uj]∈ℝ7,δvj=[Δpj,Δϕj]∈ℝ6.\hat{A}_{t:t+11}=\{\hat{a}_{j}\}_{j=t}^{t+11},\qquad\hat{a}_{j}=[\delta v_{j},u_{j}]\in\mathbb{R}^{7},\qquad\delta v_{j}=[\Delta p_{j},\Delta\phi_{j}]\in\mathbb{R}^{6}.
𝒬tAbs=(ℓ,Itwrist,Itthird,XtAbs),XtAbs={[𝒱⁡(Tb,et−i),gt−i]}i=03.\mathcal{Q}_{t}^{\mathrm{Abs}}=\left(\ell,I_{t}^{\mathrm{wrist}},I_{t}^{\mathrm{third}},X_{t}^{\mathrm{Abs}}\right),\qquad X_{t}^{\mathrm{Abs}}=\left\{\left[\mathcal{V}\!\left(T_{b,e_{t-i}}\right),g_{t-i}\right]\right\}_{i=0}^{3}.
ℛtWorld​-​Δ=A^t:t+11World​-​Δ,δvj=𝒱(T¯b,ej+1)−𝒱(Tb,ej),\mathcal{R}_{t}^{\mathrm{World}\text{-}\Delta}=\hat{A}_{t:t+11}^{\mathrm{World}\text{-}\Delta},\qquad\delta v_{j}=\mathcal{V}(\bar{T}_{b,e_{j+1}})-\mathcal{V}(T_{b,e_{j}}),
𝒱⁡(T¯b,ej+1)=𝒱⁡(Tb,ej)+δ​vj.\mathcal{V}(\bar{T}_{b,e_{j+1}})=\mathcal{V}(T_{b,e_{j}})+\delta v_{j}.
𝒬tAbs=(ℓ,Itwrist,Itthird,XtAbs),XtAbs={[𝒱⁡(Tb,et−i),gt−i]}i=03.\mathcal{Q}_{t}^{\mathrm{Abs}}=\left(\ell,I_{t}^{\mathrm{wrist}},I_{t}^{\mathrm{third}},X_{t}^{\mathrm{Abs}}\right),\qquad X_{t}^{\mathrm{Abs}}=\left\{\left[\mathcal{V}\!\left(T_{b,e_{t-i}}\right),g_{t-i}\right]\right\}_{i=0}^{3}.
ℛtEEF​-​Δ=A^t:t+11EEF​-​Δ,δvj=𝒱(ΔTj),ΔTj=Tb,ej−1T¯b,ej+1.\mathcal{R}_{t}^{\mathrm{EEF}\text{-}\Delta}=\hat{A}_{t:t+11}^{\mathrm{EEF}\text{-}\Delta},\qquad\delta v_{j}=\mathcal{V}(\Delta T_{j}),\qquad\Delta T_{j}=T_{b,e_{j}}^{-1}\bar{T}_{b,e_{j+1}}.
Δ​Tj=[Euler−1⁡(Δ​ϕj)Δ​pj01],T¯b,ej+1=Tb,ej​Δ​Tj.\Delta T_{j}=\begin{bmatrix}\operatorname{Euler}^{-1}(\Delta\phi_{j})&\Delta p_{j}\\ 0&1\end{bmatrix},\qquad\bar{T}_{b,e_{j+1}}=T_{b,e_{j}}\Delta T_{j}.
𝒬tEEF​-​Δ=(ℓ,Itwrist,Itthird,XtEEF​-​Δ),\mathcal{Q}_{t}^{\mathrm{EEF}\text{-}\Delta}=\left(\ell,I_{t}^{\mathrm{wrist}},I_{t}^{\mathrm{third}},X_{t}^{\mathrm{EEF}\text{-}\Delta}\right),
XtEEF​-​Δ={[𝒱⁡(Tb,et−1​Tb,et−i),gt−i]}i=03={[𝒱⁡(Tet,et−i),gt−i]}i=03.X_{t}^{\mathrm{EEF}\text{-}\Delta}=\left\{\left[\mathcal{V}\!\left(T_{b,e_{t}}^{-1}T_{b,e_{t-i}}\right),g_{t-i}\right]\right\}_{i=0}^{3}=\left\{\left[\mathcal{V}\!\left(T_{e_{t},e_{t-i}}\right),g_{t-i}\right]\right\}_{i=0}^{3}.
ℛtWorld​-​Δ=A^t:t+11World​-​Δ,δvj=𝒱(T¯b,ej+1)−𝒱(Tb,ej),\mathcal{R}_{t}^{\mathrm{World}\text{-}\Delta}=\hat{A}_{t:t+11}^{\mathrm{World}\text{-}\Delta},\qquad\delta v_{j}=\mathcal{V}(\bar{T}_{b,e_{j+1}})-\mathcal{V}(T_{b,e_{j}}),
𝒱⁡(T¯b,ej+1)=𝒱⁡(Tb,ej)+δ​vj.\mathcal{V}(\bar{T}_{b,e_{j+1}})=\mathcal{V}(T_{b,e_{j}})+\delta v_{j}.
𝒬tEEF​-​Δ=(ℓ,Itwrist,Itthird,XtEEF​-​Δ),\mathcal{Q}_{t}^{\mathrm{EEF}\text{-}\Delta}=\left(\ell,I_{t}^{\mathrm{wrist}},I_{t}^{\mathrm{third}},X_{t}^{\mathrm{EEF}\text{-}\Delta}\right),
XtEEF​-​Δ={[𝒱⁡(Tb,et−1​Tb,et−i),gt−i]}i=03={[𝒱⁡(Tet,et−i),gt−i]}i=03.X_{t}^{\mathrm{EEF}\text{-}\Delta}=\left\{\left[\mathcal{V}\!\left(T_{b,e_{t}}^{-1}T_{b,e_{t-i}}\right),g_{t-i}\right]\right\}_{i=0}^{3}=\left\{\left[\mathcal{V}\!\left(T_{e_{t},e_{t-i}}\right),g_{t-i}\right]\right\}_{i=0}^{3}.
ℛtEEF​-​Δ=A^t:t+11EEF​-​Δ,δvj=𝒱(ΔTj),ΔTj=Tb,ej−1T¯b,ej+1.\mathcal{R}_{t}^{\mathrm{EEF}\text{-}\Delta}=\hat{A}_{t:t+11}^{\mathrm{EEF}\text{-}\Delta},\qquad\delta v_{j}=\mathcal{V}(\Delta T_{j}),\qquad\Delta T_{j}=T_{b,e_{j}}^{-1}\bar{T}_{b,e_{j+1}}.
Δ​Tj=[Euler−1⁡(Δ​ϕj)Δ​pj01],T¯b,ej+1=Tb,ej​Δ​Tj.\Delta T_{j}=\begin{bmatrix}\operatorname{Euler}^{-1}(\Delta\phi_{j})&\Delta p_{j}\\ 0&1\end{bmatrix},\qquad\bar{T}_{b,e_{j+1}}=T_{b,e_{j}}\Delta T_{j}.

Appendix M Additional Problem-Setup Details

Demonstrations pair observations and actions for a robot/task combination; state/action encoding follows each experimental variant. The policy maps observation history plus instruction to an action chunk. The stated controls hold unless explicitly varied, particularly the growing total budget in the fixed-per-source sweep and target-data replacement in exposure studies.

Concept demonstration

Open image ↗