Literature
Display
← All papers

Contractive Dynamical Imitation Policies for Efficient Out-of-Sample Recovery

Amin Abyaneh · Mahrokh G. Boroujeni · Hsiu-Chin Lin · Giancarlo Ferrari-Trecate

· ICLR 2025 ↗

Robotics8 min text · CondensedOriginal paper ↗
Contents

At a glance

SCDS learns trajectory-generating policies from state demonstrations using contractive recurrent equilibrium networks and invertible output layers. The architecture supports ordinary gradient optimization and an adjustable contraction rate. Simulated handwriting, navigation, and manipulation tests show lower errors from perturbed starting states. The paper also presents loss bounds for a bounded initial-state region; these concern the modeled trajectories and rely on stated mathematical and controller assumptions.

1 Introduction

Asymptotic convergence can permit large transient deviations from demonstrations. SCDS instead promotes contraction between trajectories. It combines a recurrent equilibrium network (REN), a learned latent space, and coupling layers, training on complete state rollouts rather than measured velocities. The intended benefit is faster recovery toward demonstrated behavior after a change of initial state. [11], [19], [38].

Refer to caption
Figure 1: Policy rollouts generated by contractive and stable policies. While both policies eventually reach the target, the contractive policy closely mimics the expert in the transient phase.
Refer to caption
Figure 2: Overview of the SCDS training scheme. The policy structure (top box) consists of a REN, a linear projection, and a bijection block, ensuring contractivity for any choice of parameters, 𝜽\boldsymbol{\theta}. In the forward pass (→), initial states are passed through a differentiable ODE solver to generate state rollouts. The loss function penalizes the discrepancy between the generated and expert trajectories, updating the policy via backpropagation (←).

2 Contractive Policy Formulation

The policy consists of latent dynamics, an initialization map, and an output transformation.

2.1 Problem Setup

Training data and the desired recovery behavior define the learning problem.

Dataset.

Demonstrations contain sampled states with potentially different lengths and initial conditions. The formal setup assumes a common target and does not require state derivatives as training labels.

Equation 1
𝒟≜{𝐲m=(𝐲0m,⋯,𝐲Nm−1m)}m=1M,\mathcal{D}\triangleq\Bigl\{\mathbf{y}^{m}=\big(\mathbf{y}^{m}_{0},\;\cdots,\;\mathbf{y}^{m}_{N_{m}-1}\big)\Bigr\}_{m=1}^{M}, (1)

Objective.

The objective combines faithful reproduction with recovery from unseen initial states or perturbations. Stability of the trajectory generator is the proposed mechanism.

2.2 Background: Contractive latent dynamics

A continuous REN parameterization guarantees exponential convergence of latent trajectories for its allowed activation functions and constructed matrices. The rate can be prescribed or learned. Its implicit nonlinear state is required to be uniquely defined. [41].

Equations 2–3
‖𝐳a​(t)−𝐳b​(t)‖≤α​e−γ​t​‖𝐳a​(0)−𝐳b​(0)‖∀t≥0,\bigl\|\mathbf{z}^{a}(t)-\mathbf{z}^{b}(t)\bigr\|\leq\alpha\,e^{-\gamma t}\bigl\|\mathbf{z}^{a}(0)-\mathbf{z}^{b}(0)\bigr\|\quad\forall t\geq 0\,, (2)
[𝐳˙​(t)𝒗⁡(t)]=Ω⁡(𝜽,γ)​[𝐳⁡(t)σ⁡(𝒗⁡(t))],\displaystyle\begin{bmatrix}\dot{\mathbf{z}}(t)\\ \boldsymbol{v}(t)\end{bmatrix}=\Omega(\boldsymbol{\theta},\gamma)\begin{bmatrix}\mathbf{z}(t)\\ \sigma\bigl(\boldsymbol{v}(t)\bigr)\end{bmatrix}, (3)

2.3 Expressive Contractive Policies

A learned linear projection changes latent dimension, and invertible coupling layers add nonlinear output flexibility. Initialization inverts these layers and applies a pseudoinverse, potentially only approximating the requested state. The authors state that the output preserves contraction and its rate. A separate low-level controller must track the generated states on the robot.

Equations 4–7
ϕ𝜽:{𝐲^​(t)=g𝜽​(𝐳⁡(t))(output transformation)𝐳⁡(0)=h𝜽​(𝐲0)(initial condition)𝐳˙​(t)=f𝜽​(𝐳⁡(t))(latent dynamics),\displaystyle\phi_{\boldsymbol{\theta}}:\begin{cases}\hat{\mathbf{y}}(t)=g_{\boldsymbol{\theta}}\bigl(\mathbf{z}(t)\bigr)\qquad&\textit{(output transformation)}\\ \mathbf{z}(0)=h_{\boldsymbol{\theta}}\bigl(\mathbf{y}_{0}\bigr)\qquad&\textit{(initial condition)}\\ \dot{\mathbf{z}}(t)=f_{\boldsymbol{\theta}}\bigl(\mathbf{z}(t)\bigr)\qquad&\textit{(latent dynamics)}\end{cases}\quad, (4)
𝐲^(t)=g𝜽(𝐳(t))≜g𝜽,1(⋯(g𝜽,K(𝑷𝜽𝐳(t))),\displaystyle\hat{\mathbf{y}}(t)=g_{\boldsymbol{\theta}}\bigl(\mathbf{z}(t)\bigr)\triangleq g_{\boldsymbol{\theta},1}\bigl(\cdots\bigl(g_{\boldsymbol{\theta},K}\bigl(\boldsymbol{P}_{\boldsymbol{\theta}}\,\mathbf{z}(t)\bigr)\bigr), (5)
𝐳(0)=h𝜽(𝐲0)=𝑷𝜽†g𝜽,K−1(⋯g𝜽,1−1(𝐲0)).\displaystyle\mathbf{z}(0)=h_{\boldsymbol{\theta}}(\mathbf{y}_{0})=\boldsymbol{P}_{\boldsymbol{\theta}}^{\dagger}\,g_{\boldsymbol{\theta},K}^{-1}\bigl(\cdots g_{\boldsymbol{\theta},1}^{-1}\bigl(\mathbf{y}_{0}\bigr)\bigr). (6)
ϕ𝜽:{𝐲^(t)=g𝜽,1(⋯(g𝜽,K(𝑷𝜽𝐳(t)))(output map)𝐳(0)=𝑷𝜽†g𝜽,K−1(⋯g𝜽,1−1(𝐲0))(initial condition)[𝐳˙​(t)𝒗⁡(t)]=Ω⁡(𝜽,γ)​[𝐳⁡(t)σ⁡(𝒗⁡(t))](latent dynamics).\displaystyle\phi_{\boldsymbol{\theta}}:\begin{cases}\hat{\mathbf{y}}(t)=g_{\boldsymbol{\theta},1}\bigl(\cdots\bigl(g_{\boldsymbol{\theta},K}\bigl(\boldsymbol{P}_{\boldsymbol{\theta}}\,\mathbf{z}(t)\bigr)\bigr)&(\textit{output map})\\ \mathbf{z}(0)=\boldsymbol{P}_{\boldsymbol{\theta}}^{\dagger}\,g_{\boldsymbol{\theta},K}^{-1}\bigl(\cdots g_{\boldsymbol{\theta},1}^{-1}\bigl(\mathbf{y}_{0}\bigr)\bigr)&(\textit{initial condition})\\ \hskip-2.0pt\begin{bmatrix}\dot{\mathbf{z}}(t)\\ \boldsymbol{v}(t)\end{bmatrix}=\Omega(\boldsymbol{\theta},\gamma)\begin{bmatrix}\mathbf{z}(t)\\ \sigma\bigl(\boldsymbol{v}(t)\bigr)\end{bmatrix}&(\textit{latent dynamics})\end{cases}\quad. (7)

3 Learning Contractive Policies through Imitation

Training differentiates through policy rollouts and measures their discrepancy from expert state trajectories.

3.1 Differentiable ODE solutions

A differentiable ODE solver produces rollouts over a selected horizon; multiple initial conditions can be integrated in parallel. Longer horizons increase computation and memory demands.

3.2 Trajectory space loss

MSE compares aligned, equal-length trajectories, while soft-DTW accommodates temporal alignment and unequal lengths. A proposed inverse-distance weighting emphasizes demonstrations with nearby starting states. Experiments use rollouts initialized at demonstration starts and average their corresponding imitation losses. [38].

Equation 8
L⁡(𝐲^0,𝜽)≜∑m=1Mλm​(𝐲^0)​ℓ​(𝐲^,𝐲m),s.t.​𝐲^=i​v​p​(ϕ𝜽,𝐲^0),\displaystyle L(\hat{\mathbf{y}}_{0};\;\boldsymbol{\theta})\triangleq\sum_{m=1}^{M}\lambda_{m}(\hat{\mathbf{y}}_{0})\,\ell\bigl(\hat{\mathbf{y}},\;\mathbf{y}^{m}\bigr),\;\;\text{s.t.}\;\hat{\mathbf{y}}=ivp(\phi_{\boldsymbol{\theta}},\;\hat{\mathbf{y}}_{0}), (8)
Refer to caption
Figure 3: Uniform weighting averages the deviation from both demonstrations, hence, L⁡(𝐲0a,𝜽)≠0L(\mathbf{y}_{0}^{a};\;\boldsymbol{\theta})\neq 0. In contrast, Eq. 9 results in L⁡(𝐲0a,𝜽)=0L(\mathbf{y}_{0}^{a};\;\boldsymbol{\theta})=0. For 𝐲0c\mathbf{y}_{0}^{c}, the weights in Eq. 9 assign higher importance to the closer 𝐲0b\mathbf{y}_{0}^{b}’s demonstration, which is intuitive.
Equations 9–10
λm​(𝐲^0)=(‖𝐲^0−𝐲0m‖22)−1∑m′=1M(‖𝐲^0−𝐲0m′‖22)−1.\displaystyle\lambda_{m}(\hat{\mathbf{y}}_{0})=\frac{\bigl(\|\hat{\mathbf{y}}_{0}-\mathbf{y}^{m}_{0}\|_{2}^{2}\bigr)^{-1}}{\sum_{m^{\prime}=1}^{M}\bigl(\|\hat{\mathbf{y}}_{0}-\mathbf{y}^{m^{\prime}}_{0}\|_{2}^{2}\bigr)^{-1}}\;\;. (9)
ℒ⁡(ρ0,𝜽)≜𝔼𝐲^0∼ρ0L⁡(𝐲^0,𝜽)→estimationunbiasedℒ^​(𝒮0,𝜽)≜1|𝒮0|​∑s=1|𝒮0|L⁡(𝐲^0s,𝜽)\displaystyle\mathcal{L}(\rho_{0};\;\boldsymbol{\theta})\triangleq\mathop{\mathbb{E}}_{\hat{\mathbf{y}}_{0}\sim\rho_{0}}L(\hat{\mathbf{y}}_{0};\;\boldsymbol{\theta})\xrightarrow[\text{estimation}]{\text{unbiased}}\hat{\mathcal{L}}({\mathcal{S}_{0}};\;\boldsymbol{\theta})\triangleq\frac{1}{{|\mathcal{S}_{0}|}}\sum_{s=1}^{{|\mathcal{S}_{0}|}}L(\hat{\mathbf{y}}^{s}_{0};\;\boldsymbol{\theta}) (10)

3.3 Optimization problem

Gradient descent minimizes the trajectory loss without adding separate contraction constraints, because the REN parameterization supplies its latent stability structure. Algorithm 1 retains the training sequence and optional contraction-rate optimization.

Equation 11
𝜽∗≜arg⁡min𝜽∈ℝNθ​ℒ^​({𝐲0m}m=1M,𝜽),\boldsymbol{\theta}^{*}\triangleq\arg\min_{\boldsymbol{\theta}\in\mathbb{R}^{N_{\theta}}}\hat{\mathcal{L}}\bigl(\bigl\{\mathbf{y}^{m}_{0}\bigr\}_{m=1}^{M};\;\boldsymbol{\theta}\bigr), (11)
Algorithm 1 State-only contractive policy learning
1: Require: Expert data 𝒟\mathcal{D}, contraction rate γ\gamma, learning rate η\eta, horizon H{H}
2: Initialize 𝜽\boldsymbol{\theta} ⊳\triangleright Policy parameter initialization
3: while not converged do
4:    Sample 𝐲^0∼𝐲0m∈𝒟\hat{\mathbf{y}}_{0}\sim\mathbf{y}^{m}_{0}\in\mathcal{D}
5:    Initialize 𝐳⁡(0)←𝑷𝜽†​g𝜽−1​(𝐲^0)\mathbf{z}(0)\leftarrow\boldsymbol{P}_{\boldsymbol{\theta}}^{\dagger}\,g_{\boldsymbol{\theta}}^{-1}(\hat{\mathbf{y}}_{0}) ⊳\triangleright Latent initial state (Eq. 6)
6:    𝐳:={𝐳t}t=0H−1\mathbf{z}:=\{\mathbf{z}_{t}\}_{t=0}^{{H}-1} ←\leftarrow Neural_ODE(ϕ𝜽,𝐳⁡(0))\bigl(\phi_{\boldsymbol{\theta}},\;\mathbf{z}(0)\bigr) ⊳\triangleright Differentiable ODE solver (Subsec. 3.1)
7:    𝐲^:={𝐲^t}t=0H−1←g𝜽​(𝐳)\hat{\mathbf{y}}:=\{\hat{\mathbf{y}}_{t}\}_{t=0}^{{H}-1}\leftarrow g_{\boldsymbol{\theta}}(\mathbf{z}) ⊳\triangleright Output map (Eq. 5)
8:    for each 𝐲m\mathbf{y}^{m} in 𝒟\mathcal{D} do
9:     Find λm,ℓ⁡(𝐲^,𝐲m)\lambda_{m},\;\ell(\hat{\mathbf{y}},\;\mathbf{y}^{m}) ⊳\triangleright Eq. 10, Eq. 9
10:     Compute ℒ^=∑mλm​ℓ​(𝐲^,𝐲m)\hat{\mathcal{L}}=\sum_{m}\lambda_{m}\ell(\hat{\mathbf{y}},\;\mathbf{y}^{m}) ⊳\triangleright Empirical loss (Subsec. 3.2)   
11:    Update 𝜽←𝜽−η∇ℒ^\boldsymbol{\theta}\leftarrow\boldsymbol{\theta}-\eta\nabla\hat{\mathcal{L}} ⊳\triangleright Gradient descent
12: return 𝜽\boldsymbol{\theta}

4 Out-of-sample performance guarantees

The authors state worst-case and expected MSE bounds for equal-length rollouts whose initial states lie in a bounded multifocal-ellipse region around demonstration starts. The bound separates training error from an initial-state uncertainty term controlled by region size and contraction constants. It does not directly certify soft-DTW loss, task completion, collision avoidance, or low-level tracking.

Equations 12
𝒴0={𝐲^0|∑m=1M‖𝐲^0−𝐲0m‖2≤R},\displaystyle\mathcal{Y}_{0}=\Bigl\{\hat{\mathbf{y}}_{0}\Big|\sum_{m=1}^{M}\|\hat{\mathbf{y}}_{0}-\mathbf{y}^{m}_{0}\|_{2}\leq R\Bigr\}, (12)
L⁡(𝐲^0,𝜽)≤∑m=1Mλm​(𝐲^0)​ℓ​(𝐲^m,𝐲m)⏟(i)+α2​R2​(e−2​γ−1)H​M​(e−2​γH−1)⏟(ii),\displaystyle L(\hat{\mathbf{y}}_{0};\;\boldsymbol{\theta})\leq\underbrace{\sum_{m=1}^{M}\lambda_{m}(\hat{\mathbf{y}}_{0})\,\ell(\hat{\mathbf{y}}^{m},\mathbf{y}^{m})}_{\text{(i)}}+\underbrace{\frac{\alpha^{2}\,R^{2}\,(e^{-2\gamma}-1)}{H\,M\,(e^{\frac{-2\gamma}{H}}-1)}}_{\text{(ii)}},
Refer to caption
Figure 4: In-sample and OOS policy rollouts for selected 2D tasks in the LASA dataset. The training process promotes higher contraction rates ( App. E.2), resulting in effective OOS recovery.
Refer to caption
Figure 5: Comparing SCDS to the selected baselines on in-sample and OOS rollouts in the 2D task space. BC results in diverging trajectories due to the lack of stability guarantees. While SNDS and SDS-EF ensure global stability and reach the target, they display large deviations from expert data. In contrast, SCDS efficiently recovers from OOS states through its contracting transient behavior.
Equation
ℒ⁡(ρ0,𝜽)≤maxm∈{1,⋯,M}⁡ℓ⁡(𝐲^m,𝐲m)+α2​R2​(e−2​γ−1)H​M​(e−2​γH−1),\displaystyle\mathcal{L}(\rho_{0};\;\boldsymbol{\theta})\leq\max_{m\in\{1,\cdots,M\}}\ell(\hat{\mathbf{y}}^{m},\mathbf{y}^{m})+\frac{\alpha^{2}\,R^{2}\,(e^{-2\gamma}-1)}{H\,M\,(e^{\frac{-2\gamma}{H}}-1)},

5 Experiments

Experiments use LASA handwriting and Robomimic demonstrations, followed by simulated deployment in Isaac Lab.

5.1 Setup and baselines

Baselines are SNDS, SDS-EF, and ordinary behavior cloning. Out-of-sample starts are sampled near demonstrated starts, with radius 10% of their norm. MSE and soft-DTW measure trajectory error. The paper compares state-only SCDS with baselines trained using state derivatives.

Refer to caption
Figure 6: In-sample and OOS policy rollouts for the 6D Can and Lift tasks in the Robomimic dataset, using both 11 and 5050 demonstrations. SCDS generates contractive trajectories even in higher-dimensional spaces, where expert trajectories exhibit non-converging patterns.

5.2 Results and discussion

SCDS has the lowest reported out-of-sample errors across the four benchmark representations. Robomimic-14D MSE is 2.68 ± 0.65 versus the best baseline’s 6.88 ± 2.16; in-sample advantages are smaller and not universal. Some Robomimic demonstrations have different endpoints, departing from the formal common-target assumption. Reported numerical bounds use an empirically estimated contraction constant. Franka and Jackal deployment tests are simulated, with 50 starting states each.

6 Conclusion, Limitations, and Future work

Contraction-focused trajectory learning improves recovery in the reported tests, at the cost of longer training and sensitivity to horizon choices. The method assumes a suitable low-level controller. Broader robot classes, longer tasks, and stochastic policies are proposed extensions rather than demonstrated results.

Table 1: Evaluating in-sample and out-of-sample rollouts error on the LASA and the Robomimic datasets. The lowest (best) values for each metric and dataset are highlighted. The expressive power of SCDS enables it to achieve near state-of-the-art performance on in-sample data. More notably, it consistently outperforms other baselines in OOS recovery, and conforms with the theoretical upper bound for MSE loss, ℒubMSE\mathcal{L}_{\text{ub}}^{\text{MSE}}, as derived in Sec. 4.
Expert   LASA-2D LASA-4D Robomimic-6D Robomimic-14D
Metric MSE soft-DTW MSE soft-DTW MSE soft-DTW MSE soft-DTW
SNDS 0.02 ±\pm 0.01 0.72 ±\pm 0.14 0.03 ±\pm 0.01 1.04 ±\pm 0.19 0.65 ±\pm 0.55 1.26 ±\pm 0.70 2.42 ±\pm 1.37 4.15 ±\pm 0.92
BC 0.04 ±\pm 0.02 0.98 ±\pm 0.12 0.05 ±\pm 0.03 1.48 ±\pm 0.16 0.56 ±\pm 0.32 1.88 ±\pm 0.20 1.75 ±\pm 0.22 4.86 ±\pm 0.58
SDS-EF 0.03 ±\pm 0.01 0.85 ±\pm 0.13 0.05 ±\pm 0.02 1.10 ±\pm 0.15 0.50 ±\pm 0.28 1.01 ±\pm 0.66 3.30 ±\pm 0.75 5.65 ±\pm 0.58
SCDS 0.02 ±\pm 0.01 0.65 ±\pm 0.05 0.03 ±\pm 0.01 0.72 ±\pm 0.12 0.56 ±\pm 0.22 1.05 ±\pm 0.37 1.68 ±\pm 0.45 4.10 ±\pm 0.40
SNDS 2.73 ±\pm 1.67 6.91 ±\pm 1.46 3.65 ±\pm 2.12 9.85 ±\pm 0.63 1.58 ±\pm 0.93 3.27 ±\pm 1.88 6.88 ±\pm 2.16 12.17 ±\pm 2.70
BC 8.63 ±\pm 4.05 16.25 ±\pm 5.27 19.25 ±\pm 7.34 27.48 ±\pm 6.83 11.05 ±\pm 7.41 19.94 ±\pm 9.57 44.98 ±\pm 15.11 37.82 ±\pm 14.13
SDS-EF 1.78 ±\pm 0.34 7.13 ±\pm 1.51 2.33 ±\pm 0.47 10.21 ±\pm 1.98 1.15 ±\pm 0.81 2.67 ±\pm 1.40 11.10 ±\pm 2.10 10.44 ±\pm 1.66
SCDS 0.32 ±\pm 0.15 1.72 ±\pm 0.54 1.09 ±\pm 0.21 2.58 ±\pm 0.30 0.89 ±\pm 0.26 1.71 ±\pm 0.53 2.68 ±\pm 0.65 6.27 ±\pm 0.48
ℒubMSE\mathcal{L}_{\text{ub}}^{\text{MSE}} 0.49 ±\pm 0.07 1.33 ±\pm 0.14 1.43 ±\pm 0.22 2.91 ±\pm 0.45
Refer to caption
Figure 7: Simulation rollouts with Franka arm and Jackal mobile robots in Isaac Lab. Both robots start from a random OOS initial state and swiftly converge to expert data.

Acknowledgments

Support came from NCCR Automation and the NECON project through the Swiss National Science Foundation.

Appendix

The appendices retain proofs, design choices, ablations, implementation details, and datasets.

Appendix A Mathematical proofs

The authors provide derivations for output contraction and the proposed loss bounds.

A.1 Proposition 2.1

The proof combines a singular-value bound for the linear projection with a coordinate-change argument for bijective layers. Its displayed constant divides by a minimum singular value, so a nonzero lower bound is required for that argument; rank-deficient projections need separate care.

Equations 13
σm​i​n​(𝑷)2≤‖𝑷​𝒗‖22‖𝒗‖22.\sigma_{min}(\boldsymbol{P})^{2}\leq\frac{\|\boldsymbol{P}\boldsymbol{v}\|_{2}^{2}}{\|\boldsymbol{v}\|_{2}^{2}}.
‖𝑷𝜽​𝐳a​(t)−𝑷𝜽​𝐳b​(t)‖\displaystyle\bigl\|\boldsymbol{P}_{\boldsymbol{\theta}}\,\mathbf{z}^{a}(t)-\boldsymbol{P}_{\boldsymbol{\theta}}\,\mathbf{z}^{b}(t)\bigr\| ≤‖𝑷𝜽‖​‖𝐳a​(t)−𝐳b​(t)‖\displaystyle\leq\bigl\|\boldsymbol{P}_{\boldsymbol{\theta}}\bigr\|\bigl\|\mathbf{z}^{a}(t)-\mathbf{z}^{b}(t)\bigr\|\quad (norm definition)\displaystyle(\textit{norm definition})
≤α​‖𝑷𝜽‖​e−γ​t​‖𝐳a​(0)−𝐳b​(0)‖\displaystyle\leq\alpha\bigl\|\boldsymbol{P}_{\boldsymbol{\theta}}\bigr\|e^{-\gamma t}\bigl\|\mathbf{z}^{a}(0)-\mathbf{z}^{b}(0)\bigr\|\quad (𝐳​ satisfies Eq. 2)\displaystyle(\mathbf{z}\textit{ satisfies~\autoref{eq:definition_contracting_system}})
≤α​‖𝑷𝜽‖σm​i​n2​(𝑷𝜽)​e−γ​t​‖𝑷𝜽​𝐳a​(0)−𝑷𝜽​𝐳b​(0)‖,\displaystyle\leq\frac{\alpha\bigl\|\boldsymbol{P}_{\boldsymbol{\theta}}\bigr\|}{\sigma_{min}^{2}(\boldsymbol{P}_{\boldsymbol{\theta}})}e^{-\gamma t}\bigl\|\boldsymbol{P}_{\boldsymbol{\theta}}\,\mathbf{z}^{a}(0)-\boldsymbol{P}_{\boldsymbol{\theta}}\,\mathbf{z}^{b}(0)\bigr\|,\quad (Lemma A.1)\displaystyle(\textit{Lemma~\ref{lemma:courant}}) (13)

A.2 Theorem 4.1

The proposed loss proof combines initial-state weighting, an exponential trajectory bound, and a harmonic–arithmetic mean inequality. Equation 16 invokes a triangle inequality for MSE; squared Euclidean loss does not satisfy the ordinary triangle inequality in general. The displayed bound is retained as the paper’s stated result rather than an independently established certificate.

Equations 14–20
M∑m=1M1xm≤∑m=1MxmM\displaystyle\frac{M}{\sum_{m=1}^{M}\frac{1}{x_{m}}}\leq\frac{\sum_{m=1}^{M}x_{m}}{M} ⇔1≤1M2​∑m=1M(1xm)​(∑m=1Mxm)\displaystyle\iff 1\leq\frac{1}{M^{2}}\sum_{m=1}^{M}\Bigl(\frac{1}{x_{m}}\Bigr)\Bigl(\sum_{m=1}^{M}x_{m}\Bigr)
⇔1≤1M2​∑m=1M∑m′=1Mxm′xm.\displaystyle\iff 1\leq\frac{1}{M^{2}}\sum_{m=1}^{M}\sum_{m^{\prime}=1}^{M}\frac{x_{m^{\prime}}}{x_{m}}. (14)
1M2​∑m=1M∑m′=1Mxm′xm≥(∏m=1M∏m′=1Mxm′xm)1/M2=1.\displaystyle\frac{1}{M^{2}}\sum_{m=1}^{M}\sum_{m^{\prime}=1}^{M}\frac{x_{m^{\prime}}}{x_{m}}\geq(\prod_{m=1}^{M}\prod_{m^{\prime}=1}^{M}\frac{x_{m^{\prime}}}{x_{m}})^{1/{M^{2}}}=1. (15)
L⁡(𝐲^0,𝜽)\displaystyle L(\hat{\mathbf{y}}_{0};\;\boldsymbol{\theta}) =∑m=1Mλm​(𝐲^0)​ℓ​(𝐲^,𝐲m)\displaystyle=\sum_{m=1}^{M}\lambda_{m}(\hat{\mathbf{y}}_{0})\,\ell\bigl(\hat{\mathbf{y}},\;\mathbf{y}^{m}\bigr)
≤∑m=1Mλm​(𝐲^0)​ℓ​(𝐲^,𝐲^m)+∑m=1Mλm​(𝐲^0)​ℓ​(𝐲^m,𝐲m),\displaystyle\leq\sum_{m=1}^{M}\lambda_{m}(\hat{\mathbf{y}}_{0})\,\ell(\hat{\mathbf{y}},\;\hat{\mathbf{y}}^{m})+\sum_{m=1}^{M}\lambda_{m}(\hat{\mathbf{y}}_{0})\,\ell(\hat{\mathbf{y}}^{m},\mathbf{y}^{m}), (16)
ℓ⁡(𝐲^,𝐲^m)\displaystyle\ell(\hat{\mathbf{y}},\;\hat{\mathbf{y}}^{m}) =1H​∑i=0H−1‖𝐲^i−𝐲^im‖22\displaystyle=\frac{1}{H}\sum_{i=0}^{H-1}\bigl\|\hat{\mathbf{y}}_{i}-\hat{\mathbf{y}}^{m}_{i}\bigr\|_{2}^{2}
≤1H∑i=0H−1α2e−2γi/H∥𝐲^0−𝐲^m0∥22\displaystyle\leq\frac{1}{H}\sum_{i=0}^{H-1}\alpha^{2}e^{-2\gamma i/H}\bigl\|\hat{\mathbf{y}}_{0}-\hat{\mathbf{y}}^{m}_{0}\bigr\|_{2}^{2}
=α2​(e−2​γ−1)H(e−2γ/H−1)​‖𝐲^0−𝐲^0m‖22,\displaystyle=\frac{\alpha^{2}\,(e^{-2\gamma}-1)}{H\,(e^{-2\gamma/H}-1)}\bigl\|\hat{\mathbf{y}}_{0}-\hat{\mathbf{y}}^{m}_{0}\bigr\|_{2}^{2}, (17)
∑m=1Mλm​(𝐲^0)​ℓ​(𝐲^,𝐲^m)≤α2​(e−2​γ−1)H(e−2γ/H−1)⋅M∑m=1M(‖𝐲^0−𝐲0m‖22)−1.\displaystyle\sum_{m=1}^{M}\lambda_{m}(\hat{\mathbf{y}}_{0})\,\ell(\hat{\mathbf{y}},\;\hat{\mathbf{y}}^{m})\leq\frac{\alpha^{2}\,(e^{-2\gamma}-1)}{H\,(e^{-2\gamma/H}-1)}\cdot\frac{M}{\sum_{m=1}^{M}\bigl(\|\hat{\mathbf{y}}_{0}-\mathbf{y}^{m}_{0}\|_{2}^{2}\bigr)^{-1}}\;\;. (18)
M∑m=1M1‖𝐲^0−𝐲0m‖22≤∑m=1M‖𝐲^0−𝐲0m‖22M≤(∑m=1M‖𝐲^0−𝐲0m‖2)2M≤R2M,\displaystyle\frac{M}{\sum_{m=1}^{M}\frac{1}{\|\hat{\mathbf{y}}_{0}-\mathbf{y}^{m}_{0}\|_{2}^{2}}}\leq\frac{\sum_{m=1}^{M}\|\hat{\mathbf{y}}_{0}-\mathbf{y}^{m}_{0}\|_{2}^{2}}{M}\leq\frac{\Bigl(\sum_{m=1}^{M}\|\hat{\mathbf{y}}_{0}-\mathbf{y}^{m}_{0}\|_{2}\Bigr)^{2}}{M}\leq\frac{R^{2}}{M}, (19)
L⁡(𝐲^0,𝜽)≤α2​(e−2​γ−1)H(e−2γ/H−1)⋅R2M+∑m=1Mλm​(𝐲^0)​ℓ​(𝐲^m,𝐲m).\displaystyle L(\hat{\mathbf{y}}_{0};\;\boldsymbol{\theta})\leq\frac{\alpha^{2}\,(e^{-2\gamma}-1)}{H\,(e^{-2\gamma/H}-1)}\cdot\frac{R^{2}}{M}+\sum_{m=1}^{M}\lambda_{m}(\hat{\mathbf{y}}_{0})\,\ell(\hat{\mathbf{y}}^{m},\mathbf{y}^{m}). (20)

A.3 Corollary 4.1.1

The expected-loss result follows from the stated pointwise bound by replacing a weighted training error with its maximum and taking expectation over the allowed initial-state region.

Equations 21
L⁡(𝐲^0,𝜽)≤maxm∈{1,⋯,M}⁡ℓ⁡(𝐲^m,𝐲m)+α2​R2​(e−2​γ−1)HM(e−2γ/H−1),\displaystyle L(\hat{\mathbf{y}}_{0};\;\boldsymbol{\theta})\leq\max_{m\in\{1,\cdots,M\}}\ell(\hat{\mathbf{y}}^{m},\mathbf{y}^{m})+\frac{\alpha^{2}\,R^{2}\,(e^{-2\gamma}-1)}{H\,M\,(e^{-2\gamma/H}-1)},
ℒ⁡(ρ0,𝜽)=𝔼𝐲^0∼ρ0L⁡(𝐲^0,𝜽)≤maxm∈{1,⋯,M}⁡ℓ⁡(𝐲^m,𝐲m)+α2​R2​(e−2​γ−1)HM(e−2γ/H−1).\displaystyle\mathcal{L}(\rho_{0};\;\boldsymbol{\theta})=\mathop{\mathbb{E}}_{\hat{\mathbf{y}}_{0}\sim\rho_{0}}L(\hat{\mathbf{y}}_{0};\;\boldsymbol{\theta})\leq\max_{m\in\{1,\cdots,M\}}\ell(\hat{\mathbf{y}}^{m},\mathbf{y}^{m})+\frac{\alpha^{2}\,R^{2}\,(e^{-2\gamma}-1)}{H\,M\,(e^{-2\gamma/H}-1)}. (21)

Appendix B Extended related work

The related-work comparison separates constrained contraction methods from architectures with built-in structure.

Table 2: Comparing contractive DS methods for imitation learning. Except for NCDS and SCDS, the rest attempt to solve a computationally challenging constrained optimization problem.
Unconstrained optimization Adjustable contraction rate State-only Learn in latent space
C-GMR (Blocher et al., 2017) X X X X
CDSP (Ravichandar et al., 2017) X X X X
CVF (Sindhwani et al., 2018) X X X X
NCDS (Mohammadi et al., 2024) ✓\checkmark X X only lower-dimensional
SCDS (ours) ✓\checkmark ✓\checkmark ✓\checkmark ✓\checkmark

B.1 Stable dynamical policies with restricted representations

Earlier Lyapunov-based approaches restrict either the motion field or certificate to make optimization tractable. Learning more flexible certificates can improve expressiveness but complicates joint optimization.

B.2 Stable neural dynamical policies

Neural stable policies use coordinate transformations or projection-based enforcement. The authors contrast their contraction objective with these asymptotic-stability approaches and their representation or smoothness limitations.

B.3 Imitation policies with no stability guarantees

Adversarial imitation, noise injection, and dataset aggregation address imitation and distribution shift through different mechanisms. They do not provide the particular contraction certificate targeted here.

B.4 Recurrent equilibrium networks (RENs)

RENs supply a flexible family of implicit dynamical models with structured stability properties. SCDS uses the continuous-time formulation, with a discrete variant also implemented.

Appendix C Design considerations

Several choices trade imitation accuracy against recovery speed and computational cost.

C.1 Contraction rate

A larger contraction rate accelerates convergence between rollouts but restricts expressiveness. Nonlinear output layers help recover modeling flexibility.

C.2 Horizon

Longer integration horizons resolve intricate or longer motions at greater training cost. Solver choice and numerical tolerance also affect this trade-off.

Equation 22
𝐳t+1=𝐳t+f𝜽(𝐳t)Δt,∀t∈0,1,…,H.\displaystyle\mathbf{z}_{t+1}=\mathbf{z}_{t}+f_{\boldsymbol{\theta}}(\mathbf{z}_{t})\Delta t,\quad\forall t\in{0,1,\ldots,{H}}. (22)

C.3 Soft DTW

Soft-DTW replaces the discrete minimum over alignments with a differentiable soft minimum. Its smoothing parameter controls approximation to ordinary DTW.

Equation 23
soft-​D​T​Wβ​(𝐲^,𝐲m)=minτ∈𝒜⁡(𝐲^,𝐲m)⁡∑(i,j)∈τβ⁡d​(y^i,yjm)2,\text{soft-}DTW^{\beta}(\hat{\mathbf{y}},\mathbf{y}^{m})=\min_{\tau\in\mathcal{A}(\hat{\mathbf{y}},\mathbf{y}^{m})}{}^{\beta}\sum_{(i,j)\in\tau}d(\hat{y}_{i},y^{m}_{j})^{2}, (23)

C.4 Implicit depth and invertible layers

Increasing the REN implicit-state dimension adds capacity but can be expensive. The authors favor additional invertible output layers for a more efficient expressiveness increase.

Appendix D Continuous REN properties

The appendix summarizes prior continuous-REN results used by the architecture.

D.1 Detailed structure

The general REN includes state, input, output, and implicit nonlinear variables. SCDS removes external inputs and biases and replaces the original output equation with its learned projection and coupling layers.

Equations 24–25
[x˙​(t)v⁡(t)y⁡(t)]=[AB1B2C1D11D12C2D21D22]⏟A~​[x⁡(t)w⁡(t)u⁡(t)]+[bxbvby]⏟b~,\begin{bmatrix}\dot{x}(t)\\ v(t)\\ y(t)\end{bmatrix}=\underbrace{\begin{bmatrix}A&B_{1}&B_{2}\\ C_{1}&D_{11}&D_{12}\\ C_{2}&D_{21}&D_{22}\end{bmatrix}}_{\tilde{A}}\begin{bmatrix}x(t)\\ w(t)\\ u(t)\end{bmatrix}+\underbrace{\begin{bmatrix}b_{x}\\ b_{v}\\ b_{y}\end{bmatrix}}_{\tilde{b}}, (24)
w⁡(t)=σ⁡(v⁡(t)),w(t)=\sigma(v(t)), (25)
[𝐳˙​(t)𝒗⁡(t)]=Ω⁡(𝜽,γ)​[𝐳⁡(t)σ⁡(𝒗⁡(t))].\displaystyle\begin{bmatrix}\dot{\mathbf{z}}(t)\\ \boldsymbol{v}(t)\end{bmatrix}=\Omega(\boldsymbol{\theta},\gamma)\begin{bmatrix}\mathbf{z}(t)\\ \sigma\bigl(\boldsymbol{v}(t)\bigr)\end{bmatrix}.

D.2 Universal contraction

Positive-definite matrix conditions provide the cited sufficient condition for contraction of the latent REN.

Equations
[−A⊤​P−P​A−C1⊤​Λ−P​B1∗W]≻0,\begin{bmatrix}-A^{\top}P-PA&-C_{1}^{\top}\Lambda-PB_{1}\\ *&W\end{bmatrix}\succ 0,
W=2​Λ−Λ​D11−D11⊤​Λ.W=2\Lambda-\Lambda D_{11}-D_{11}^{\top}\Lambda.

D.3 Free parameterization and constructability

A free parameterization constructs the needed positive-definite matrices from Gram matrices and diagonal margins, allowing unconstrained optimization of the underlying parameters.

Equations
(−Y⊤−Y−U−Z∗W)=X⊤​X+ϵ​I,P=XP⊤​XP+ϵP​I.\displaystyle\begin{pmatrix}-Y^{\top}-Y&-U-Z\\ *&W\end{pmatrix}=X^{\top}X+\epsilon I,\quad P=X^{\top}_{P}X_{P}+\epsilon_{P}I.
[𝐳˙​(t)𝒗⁡(t)]=Ω⁡(𝜽,γ)​[𝐳⁡(t)σ⁡(𝒗⁡(t))],\displaystyle\begin{bmatrix}\dot{\mathbf{z}}(t)\\ \boldsymbol{v}(t)\end{bmatrix}=\Omega(\boldsymbol{\theta},\gamma)\begin{bmatrix}\mathbf{z}(t)\\ \sigma\bigl(\boldsymbol{v}(t)\bigr)\end{bmatrix},

Appendix E Additional experiments

Additional experiments isolate contraction rate, representation size, output layers, and computation.

E.1 Further experiments on the LASA dataset

Latent trajectories converge faster at higher rates. When the LASA start-region radius increases from 10% to 25% of the demonstrated initial-state norm, SCDS retains the lowest reported errors: MSE 0.57 in 2D and 1.44 in 4D.

Refer to caption
Figure 8: Contractive behavior of the latent state. The figure illustrates the rollouts over time for varying dimensions of the latent state, under different contraction rates. Higher contraction rates accelerate convergence in both the latent state space and the robot’s state space. The dimension of the latent state here is set to 10.
Refer to caption
Figure 9: SCDS is tested over additional motions in the LASA dataset. In all cases, the induced trajectories accurately imitate the expert, even for out-of-sample initial states.
Refer to caption
Figure 10: Out-of-sample recovery for initial states sampled from regions with different volumes according to the definition of RR in Assumption 4.1. Notably, SCDS recovers even from more distant initial states.
Table 3: Evaluating out-of-sample rollouts error on the LASA dataset. Compared to Table 1, the OOS initial states are sampled from a wider distribution with a radius of 0.25​‖𝐲0m‖20.25\|\mathbf{y}_{0}^{m}\|_{2} for the LASA dataset. The lowest (best) values for each metric and dataset are highlighted. These results confirm that the recovery behavior observed in Fig. 10 persists across all the other tasks in the LASA dataset.
Expert   LASA-2D LASA-4D
Metric MSE soft-DTW MSE soft-DTW
SNDS 4.64 ±\pm 2.50 11.06 ±\pm 2.35 6.57 ±\pm 3.56 14.78 ±\pm 1.08
BC 13.81 ±\pm 6.08 29.25 ±\pm 8.45 32.73 ±\pm 11.45 41.22 ±\pm 11.51
SDS-EF 2.67 ±\pm 0.55 12.12 ±\pm 2.29 4.19 ±\pm 0.78 16.34 ±\pm 3.28
SCDS 0.57 ±\pm 0.23 2.12 ±\pm 0.70 1.44 ±\pm 0.35 3.40 ±\pm 0.56

E.2 Learning the contraction rate

An augmented objective is used to encourage larger learned contraction rates. The plotted examples recover more quickly than policies trained at a low fixed rate.

Equation 26
ℒ^aug​({𝐲0m}m=1M,𝜽,γ)≜ℒ^​({𝐲0m}m=1M,𝜽)−μ⁡(h⁡(γ)−c),\displaystyle\hat{\mathcal{L}}_{\text{aug}}\bigl(\bigl\{\mathbf{y}^{m}_{0}\bigr\}_{m=1}^{M};\boldsymbol{\theta},\gamma\bigr)\triangleq\hat{\mathcal{L}}\bigl(\bigl\{\mathbf{y}^{m}_{0}\bigr\}_{m=1}^{M};\boldsymbol{\theta}\bigr)-\mu\left(h(\gamma)-c\right), (26)
Refer to caption
Figure 11: Policies trained to have low (first row) and high (second row) contraction rates. SCDS can promote higher contraction rates during training to achieve more robustness to out-of-sample initial states and perturbations.

E.3 Latent state dimension

Larger latent states improve imitation of the selected complex handwriting motion. A small latent model can remain stable while fitting demonstrations poorly.

Refer to caption
Figure 12: Training results for a sample motion in the LASA dataset are plotted for various latent space dimensions. Training in higher dimensional latent space is more effective, especially when facing complex expert behavior that involves multiple demonstrations. Note that the time horizon is fixed here for all the rollouts.

E.4 Ablation study: bijective layers

Removing bijective output layers worsens reconstruction. Increasing REN complexity without those layers can recover comparable accuracy, but trains more slowly in the reported comparison.

Table 4: Ablation study on invertible layers to improve the policy’s expressive power. SCDS is the standard policy formulation presented in Eq. 7. SCDS-NB uses the same REN structure but does not have bijection layers. SCDS-NB∗ makes SCDS-NB more complex by increasing the complexity of the implicit layer in the REN architecture. Still SCDS-NB∗ uses no invertible layers. The best results are displayed for 50 trials on in-sample initial states.
Expert   LASA-2D LASA-4D Robomimic-6D Robomimic-14D
Metric MSE soft-DTW MSE soft-DTW MSE soft-DTW MSE soft-DTW
SCDS 0.02 ±\pm 0.01 0.65 ±\pm 0.05 0.03 ±\pm 0.01 0.72 ±\pm 0.12 0.56 ±\pm 0.22 1.05 ±\pm 0.37 1.68 ±\pm 0.45 4.10 ±\pm 0.40
SCDS-NB 0.43 ±\pm 0.18 2.41 ±\pm 0.75 1.45 ±\pm 0.30 3.38 ±\pm 0.42 1.24 ±\pm 0.35 2.23 ±\pm 0.74 3.49 ±\pm 0.85 8.24 ±\pm 0.64
SCDS-NB∗ 0.01 ±\pm 0.00 0.50 ±\pm 0.06 0.04 ±\pm 0.02 0.79 ±\pm 0.13 0.45 ±\pm 0.15 0.88 ±\pm 0.26 1.81 ±\pm 0.24 4.87 ±\pm 0.32

E.5 Ablation study: replacing REN with a non-contractive model

Replacing the contractive REN with an unconstrained RNN produces diverging out-of-sample trajectories in the illustrated tasks, supporting the role of the latent contraction structure.

Refer to caption
Figure 13: The effect of replacing the contractive REN with a non-contractive model, specifically a recurrent neural network, is investigated across two candidate tasks. Some of the generated trajectories diverge in the no contraction case, which lacks guarantees on out-of-sample recovery.
Refer to caption
Figure 14: Distance between induced trajectories for the policies using a contractive REN (left) and a non-contractive RNN (right), with a uniformly distributed set of initial states around the true initial conditions from the LASA dataset. Since the policy in the right figure is not contractive or stable, the trajectories begin to diverge, causing the distance between neighboring trajectories to increase.

E.6 Computation time

At equal epoch and data counts, continuous SCDS trains substantially more slowly than the baselines: about 1,795 seconds for LASA-2D and 2,711 seconds for Robomimic-14D. Discrete SCDS is faster but uses more memory. The authors note that equal epochs do not imply equal convergence progress.

Table 5: Evaluating the training time on the LASA and the Robomimic datasets in seconds. SCDS using either discrete or continuous REN architecture appears to be less computationally efficient in lower dimensions, but will shorten the gap in efficiency in larger datasets. Note that even though the same number of epochs are used here, the continuous SCDS method typically requires less than half of the number of epochs to achieve the same performance. Although SCDS has longer training times, it offers the stronger contractivity guarantee compared to the stability guarantees of the stable baselines or the lack of any guarantees in the case of BC.
Expert   LASA-2D LASA-4D Robomimic-6D Robomimic-14D
SNDS 156.8 ±\pm 14.1 372.5 ±\pm 22.8 435.2 ±\pm 17.9 882.2 ±\pm 45.7
BC 57.2 ±\pm 10.1 62.9 ±\pm 12.5 91.5 ±\pm 17.4 103.0 ±\pm 36.5
SDS-EF 417.1 ±\pm 18.7 658.8 ±\pm 27.2 825.7 ±\pm 59.9 1150.2 ±\pm 94.5
SCDS-Continuous 1794.7 ±\pm 92.8 1767.2 ±\pm 94.6 2190.2 ±\pm 43.2 2711.1 ±\pm 173.5
SCDS-Discrete 381.9 ±\pm 32.6 509.0 ±\pm 18.2 610.5 ±\pm 36.3 857.4 ±\pm 64.6

E.7 Longer horizon motions

On a longer Snake motion, reported DTW errors are 0.074 at a low rate and 0.097 at a higher rate, compared with approximately 0.158 reported for SNDS. This illustrates a recovery-speed versus imitation-accuracy trade-off.

Refer to caption
Figure 15: SCDS results for the Snake dataset demonstrate its ability to handle a more complex motion compared to the tasks in the LASA dataset. SCDS is applied starting from closer (left plot) and further (middle plot) initial conditions and is compared against SNDS (right plot). The complexity of the Snake task arises from the Snake dataset’s longer time horizon and the need for rapid adaptation in DS learning.

Appendix F Implementation details

The implementation uses PyTorch and a trajectory-generation interface to robot controllers. The deployment diagram includes real-world possibilities, while the reported robot experiments are simulations.

Refer to caption
Figure 16: Simulation and real-world deployment pipelines. The trained policy can be deployed to both real and simulated robots with minimal effort. Global stability and highly contractive behavior of induced trajectories ensure safety and precision during these zero-shot transfers.

F.1 Key components of SCDS

The main components are the REN, invertible coupling layers, and differentiable trajectory loss.

Refer to caption
Figure 17: Left: Invertible coupling layer structure implemented as a part of RealNVP. The affine transformation provides increased nonlinearity while maintaining bijection, and as a result, contraction properties. Right: DTW still effectively compares trajectories with different lengths. This is particularly useful when learning long-horizon trajectories, or dealing with higher sampling rates.

(1) Continuous and discrete RENs

Continuous RENs support ODE integration at arbitrary times and adjoint-based gradients. Discrete RENs avoid an ODE solver and can be faster, with a gradient graph that grows with horizon length.

(2) Invertible coupling layers

RealNVP coupling layers keep one input partition fixed while using it to parameterize an invertible affine transformation of the other.

(3) Soft and differentiable dynamic time warping

Soft-DTW aligns trajectories with different speeds, lengths, or sample rates, providing a differentiable alternative to directly aligned MSE.

F.2 Baselines

The authors reuse and tune existing baseline implementations. They do not compare directly with NCDS, citing the absence of an official implementation at the time.

SDS-EF

SDS-EF learns a diffeomorphic mapping to a simple stable latent flow and maps trajectories back to the original state space.

SNDS

SNDS jointly learns a neural policy and Lyapunov candidate, targeting global asymptotic stability.

BC

The BC baseline is a feedforward or residual state-to-action model rather than recurrent BC. Tables 6 and 7 retain code sources, architectures, and SCDS parameter ranges, including 32–64 latent dimensions and 20–50 rollout steps.

Table 6: Baseline implementation, architecture, and hyperparameters
Baseline implementations
Method Codebase Stability certificate
SNDS (Abyaneh et al., 2024) github.com/aminabyaneh/stable-imitation-policy Lyapunov, exponential, asymptotic
SDS-EF (Rana et al., 2020) github.com/mrana6/euclideanizing_flows Lyapunov, asymptotic
BC (Pomerleau, 1988; Mandlekar et al., 2020) github.com/montaserFath/BCO None
Baseline parameters
SNDS SDS-EF BC
policy network: [2, 256, 256, 256, 2] bijection blocks: 10 network: [2, 256, 256, 256, 256, 2]
ICNN network: [2, 64, 64, 1] hidden layers: 100 type: ResNet or FeedForward
activation: nn.SoftPlus activation: nn.ELU activation: nn.ReLU
α=0.01\alpha=0.01 coupling layer: ’rfnn’ regularization: 1​e−51e-5
ϵ=0.01\epsilon=0.01 ϵ=1​e−5\epsilon=1e-5
Table 7: Hyperparameters of SCDS and their optimal values or ranges are shown for every key model parameter.
Param Description Optimal value(s)
γ\gamma Lower bound on the contraction rate of the policy [1.0,18.6][1.0,18.6], learnable
KK Number of coupling layer to increase the output nonlinearity 4​to​  104\;\;\text{to}\;\;10
N𝐳N_{\mathbf{z}} Dimension of the latent state space 32​to​  6432\;\;\text{to}\;\;64
H{H} Forward simulation horizon for the generated trajectory 20​to​  5020\;\;\text{to}\;\;50
KK Number of invertible layers in the output map (also used for hidden blocks) 4,  84,\;\;8
β\beta Complexity-accuracy trade-off parameter for soft-DTW (γ\gamma in some references) 0.10.1

F.3 Computational resources

Training uses an RTX 4090 GPU and a machine with 64 GB RAM. Isaac Lab simulations run remotely and exploit parallel environments.

F.4 Simulation in Isaac Lab

Isaac Lab evaluates randomized robot initial conditions in parallel. Generated states are converted to the required joint or task-space commands; the reported benchmark evaluation uses 128 sampled starts.

Refer to caption
Figure 18: Efficient parallel testing of out-of-sample behavior in the Isaac Lab robotics simulator. Each robot is given a random initial condition drawn from a uniform distribution around true initial states. The parallel testing of 128 sampled initial conditions yields the results in Tab. 1.

F.5 Low-level controller

The formulation assumes a low-level controller capable of tracking the generated state sequence. Joint-position, inverse-dynamics, or inverse-kinematics controllers supply the connection to motor commands; the trajectory-generator analysis alone does not establish closed-loop tracking accuracy.

Appendix G Datasets

The datasets cover low-dimensional handwriting and higher-dimensional robot configurations.

G.1 LASA dataset

LASA provides seven demonstrations per handwriting motion with 1,000 samples each and common endpoints. Soft-DTW allows much shorter generated rollouts during fitting.

Refer to caption
Figure 19: Expert demonstrations in the LASA dataset.

G.2 Robomimic dataset

Robomimic experiments use Lift, Can, Square, and Transport demonstrations. End-effector position and orientation give six-dimensional states; seven joint positions and velocities give fourteen. The authors generally train on 80–90% of the 200–300 demonstrations per task and reserve the remainder for evaluation.

Refer to caption
Figure 20: Position and orientation data for different tasks in the Robomimic dataset.
Refer to caption
Figure 21: Data collection setup for Lift, Can, Square, and transport tasks (Mandlekar et al., 2021). We employ this collected data, which is available in the Robomimic repository.
Concept demonstration

Open image ↗