RESTA: Time-Scaling Augmentation for Data-Efficient Robot Imitation Learning

Abstract

Robot manipulation policies trained through imitation learning commonly rely on costly demonstrations recorded in one task direction and at one temporal resolution. We propose Reversal and Stride-based Trajectory Augmentation (RESTA), which exploits time-scaling data augmentation to improve data efficiency through trajectory reversal and temporal resampling. Trajectory-reversal augmentation reverses and relabels demonstrations between paired tasks with opposite objectives, reusing recorded rollouts without additional environment interaction. Because contact-rich manipulation is not exactly reversible, we treat reversal as an approximate paired-task prior rather than a physical symmetry. Temporal-stride augmentation varies observation and action sampling strides and conditions the policy on the resulting offsets, enabling fixed or randomized temporal schedules. Experiments across 16 paired bimanual manipulation tasks show that trajectory reversal is particularly effective in the low-data regime and that temporal-stride augmentation improves success under randomized temporal schedules. With 25 and 50 independently collected demonstrations per task, RESTA achieves 78.4% and 85.2% average success, respectively, matching or exceeding baselines trained with twice as many demonstrations (77.3% and 85.4%). These results show that RESTA can improve data efficiency in robot imitation learning.

1 Introduction

Imitation learning offers a direct way to teach robots manipulation skills (Zhao et al., 2023; Chi et al., 2023; Ze et al., 2024), but requires demonstrations that are costly to collect: each trajectory needs resetting the scene, executing the task, and checking the result. Since a demonstration also carries temporal structure (i.e., time direction and interval between observations and actions), a set of demonstration may thus provide potential supervision both across task directions (e.g., open vs close door) and across temporal resolutions (e.g., accelerating vs decelerating). This motivates our research question: Can temporal transformations of existing demonstrations improve imitation learning across task directions and execution schedules, without additional data collection?

We use time scaling to describe transformations of the temporal coordinate used to construct policy-training windows. In physics, mathematics, and computer simulation, time scaling refers to changing the ratio between simulation time and real-world time: it can speed up or slow down a process, or formally run its temporal coordinate backward. In our setting, the two relevant properties are time reversal and frequency change. Time reversal (Jiang et al., 2025; Qiao et al., 2026; Kim et al., 2026; Barkley et al., 2023) changes the direction of a sequence: for paired tasks such as opening and closing an articulated object, a recorded opening trajectory can provide potential supervision for closing when its observations are reversed and actions are relabeled toward the next robot configuration. This correspondence is approximate and depends on whether the reversed transitions are realizable in the environment. Frequency change, in contrast, preserves temporal order while changing the time intervals between observations or action targets. It can be achieved in several ways, including changing action persistence (Hu et al., 2024), using multi-frequency action chunks (Zhang et al., 2026), or sparsely sampling target sequences (Bai et al., 2026; Park et al., 2025). In this work, we use positive temporal strides to change frequency: sampling a trajectory every step defines the reference speed; increasing the stride samples more widely separated states and corresponds intuitively to accelerating the robot’s motion, whereas decreasing it samples more densely and corresponds to decelerating it. Thus, reversal changes the task direction of a demonstration, while our stride-based augmentation changes its temporal resolution. During training, both can be naturally jointly applied.

To exploit these two properties, we propose Reversal and Stride-based Trajectory Augmentation (RESTA), which combines paired-task trajectory reversal with temporal-stride augmentation and temporal offset conditioning. In more details, reversal uses a trajectory from one task direction to provide additional supervision for its paired direction, reducing the need to collect independent demonstrations for both. We form these examples by reversing the observation sequence and relabeling each action toward the next robot configuration. Temporal-stride augmentation samples histories and action chunks at different positive strides, allowing the same demonstrations to represent varied control frequencies. Because the policy must distinguish these frequencies, we provide the temporal offset, which contains both observation and action offsets, through a lightweight embedding. This conditioning can be adapted to different policy architectures; at inference, different offset schedules set the time intervals between predicted targets, allowing the policy to operate at higher or lower effective frequencies.

Contributions

Our contributions can be summarized as follows (Figure 1):

1. We formulate the imitation-learning training-window problem and its time-scaling view, with each window pairing an observation history with future action targets. This gives a common view of trajectory reversal and positive temporal resampling, exposing two ways to reuse temporal structure in demonstrations (Sections 4.1 and 4.2).
2. We introduce paired-task trajectory reversal as a data augmentation (Section 4.3) that creates candidate behavior-cloning examples by reversing observations and relabeling absolute next-configuration targets, without additional environment interaction.
3. We introduce temporal-stride augmentation with observation and action offset conditioning (Section 4.4) to vary the frequency in demonstration windows and tell the policy at which time intervals it should represent and predict.
4. We evaluate RESTA on sixteen paired bimanual tasks (Section 5.2). At matched data budgets, it achieves higher average success and requires fewer simulator physics steps per successful episode than the baseline, indicating more reliable task completion and shorter simulated execution time. We also compare against larger-data baselines and analyze the contribution of each augmentation.

2 Related Work

We organize the related work around two directions: action-chunk imitation learning and temporal transformations. Within temporal transformations, we distinguish time-reversal methods from approaches that change action frequency. Due to page limit, discussions of action-chunk imitation policies and temporal-frequency methods appear in Appendix A.

Temporal Symmetry for Robotics Control

Time Reversal Deep Reinforcement Learning (TR-DRL) (Jiang et al., 2025) filters dynamics-consistent reversible transitions, augments replay data, and shapes rewards using reversal; Time-Symmetric Data Augmentation (Barkley et al., 2023) also constructs reverse-time transitions for off-policy RL and studies failures from irreversible dynamics. Other uses include symmetry-guided quadruped gait transitions (Ding et al., 2024), self-supervised manipulation planning (Nair et al., 2020), and reverse curricula that alter the state distribution (Florensa et al., 2017; Tao et al., 2024). We instead construct behavior-cloning examples from paired manipulation demonstrations. Recent work also studies demonstration reversal. Reverse Playback Motion (Kim et al., 2026) has an operator execute a motion from its goal state, then reconstructs a forward trajectory by reversing recorded end-effector poses. We reverse observations and relabel targets in an existing expert dataset, without additional execution or collection. Reverse to Advance (Qiao et al., 2026) iteratively uses environment-collected easy- and hard-task rollouts, retaining segments with kinematic and critic-guided advantage filters; a learned dynamics filter appears only in a comparison baseline. Our method uses fixed expert trajectories without policy rollouts, resets, or new data, and also augments temporal strides and observation offsets.

3 Background

The background covers the foundations of action-chunk imitation learning: the Markov decision-process formulation and offline imitation learning (Sections 3.1 and 3.2), followed by 3D Diffusion Policy (DP3) as the action-chunk policy backbone (Section 3.3).

3.1 Markov Decision Processes

We use a finite-horizon Markov decision process (MDP) to model robot manipulation. It is specified by \mathcal{M}=(\mathcal{S},\mathcal{A},P,\rho,r,T), with full robot–environment state space \mathcal{S}, action space \mathcal{A}, transition kernel P, initial-state distribution \rho, reward r, and horizon T. At the start of each episode, the initial state is sampled as s_{0}\sim\rho, where \rho is the distribution over robot–environment states induced by the reset procedure. At step t, state s_{t}\in\mathcal{S} evolves as s_{t+1}\sim P(\cdot\mid s_{t},a_{t}) under action a_{t}\in\mathcal{A}. Actions specify absolute joint-position and gripper targets. A supplied low-level controller converts each target into the robot commands executed over the next transition, so its tracking behavior is included in the transition kernel P. The objective is to reach the success set \mathcal{G}\subseteq\mathcal{S} within T steps. Episodes terminate at first success or the horizon, with reward one for success and zero otherwise (provided by reward function r). The policy uses sensory observations o_{t} in a length-L history h_{t}=(o_{t-L+1},\ldots,o_{t}) instead of full states.

3.2 Imitation Learning

Offline imitation learning fits a policy to N expert trajectories without rewards or a transition model. Let \mathcal{D}=\{\tau^{(n)}\}_{n=1}^{N}, where \tau^{(n)}=(o_{0},a_{0},\ldots,a_{T_{n}-1},o_{T_{n}}) has T_{n} transitions. At timestep t, the policy receives the history h_{t}=(o_{t-L+1},\ldots,o_{t}) and predicts an H-action chunk. A training window is the pair (h_{t},A_{t}^{*}), where A_{t}^{*}=(a_{t}^{*},\ldots,a_{t+H-1}^{*}); valid windows satisfy t\geq L-1 and t+H-1<T_{n}.

At deployment, the history has length L; before L observations are available, the earliest observation is repeated to left-pad it, and then the most recent L observations are retained. The policy executes part of each predicted chunk, collects new observations, and replans. Behavior cloning (BC) fits the policy to these expert windows with the objective

\mathcal{L}_{\mathrm{IL}}(\theta)=\mathbb{E}_{(h_{t},A_{t}^{*})\sim\mathcal{D}}\left[\ell\!\left(\pi_{\theta}(\cdot\mid h_{t}),A_{t}^{*}\right)\right],

where \pi_{\theta}(\cdot\mid h_{t}) is the predictive distribution and \ell measures its disagreement with the expert chunk and the expectation ranges over valid windows. Here, windows use consecutive trajectory samples; our method later varies their temporal offsets in Section 4.2.

3.3 3D Diffusion Policy (DP3)

DP3 (Ze et al., 2024) represents \pi_{\theta}(A_{t}\mid h_{t}) through conditional denoising. An encoder e_{\phi} with parameters \phi combines point-cloud features and proprioception across h_{t} into c_{t}=e_{\phi}(h_{t}). In our action-chunk setting, the policy predicts H absolute joint-position and gripper targets from this observation history. Let \psi(o_{t}) extract the recorded joint positions and gripper configuration in the controller’s target representation. For consecutive samples, the supervised action chunk is (\psi(o_{t+1}),\ldots,\psi(o_{t+H})), with t+H\leq T. A supplied low-level controller drives the robot toward each predicted target; it is not learned, and its tracking behavior is included in the transition model. The objective is to learn action-chunk predictions from demonstrations that enable successful closed-loop task execution. For a normalized expert chunk x_{0}, diffusion step n produces

x_{n}=\sqrt{\bar{\alpha}_{n}}\,x_{0}+\sqrt{1-\bar{\alpha}_{n}}\,\epsilon,\qquad\epsilon\sim\mathcal{N}(0,I),

where x_{n} is the noisy chunk, \bar{\alpha}_{n} is cumulative signal retention, \epsilon is Gaussian noise, and I is the identity covariance. Training learns denoising conditioned on c_{t} and n using the objective

\mathcal{L}_{\mathrm{DP3}}(\theta;x_{0},c_{t})=\mathbb{E}_{n,\epsilon}\left[\left\|\epsilon-\epsilon_{\theta}(x_{n},c_{t},n)\right\|_{2}^{2}\right],

where \epsilon_{\theta} predicts the noise added at diffusion step n. Inference iteratively converts noise into an action chunk.

4 RESTA

We learn action-chunk policies from expert demonstrations using two augmentations based on the time-scaling framework introduced below. After defining the control problem and paired-task setting in Section 4.1, we introduce the time-scaling framework in Section 4.2, then construct reversed demonstrations to share supervision between task directions in Section 4.3. We then vary the temporal offset of training windows and condition the policy on that offset in Section 4.4. Section 4.5 brings these steps together for training and inference.

4.1 Problem Formulation

We consider learning robot control from a fixed dataset \mathcal{D} of expert demonstrations. The action-chunk policy and its target representation are defined in the DP3 background; here we specify the paired-task setting used for trajectory reversal. The tasks are organized into pairs with expert datasets \mathcal{D}^{+} and \mathcal{D}^{-}, such as opening and closing an articulated object. We construct reverse-task supervision by reversing recorded trajectories and relabeling each action toward the next configuration in the reversed sequence. This is an offline augmentation procedure, not an assertion that every reversed transition is physically realizable.

4.2 Time Scaling

Time scaling changes how a demonstrated trajectory is indexed to construct a training window (h_{t},A_{t}), a finite portion of the trajectory used to form one training example. These windows appear in Equation 1. Let \xi\in[0,T] denote the source coordinate along the original trajectory, with T its final source-coordinate value. For a window with transformed sampling instants 0=v_{0}<v_{1}<\cdots<v_{N}=\mathcal{V}, where \mathcal{V} is the transformed window endpoint, define the mapping \xi(v) from transformed position to source-coordinate location by

\xi(v_{0})=\xi_{0},\qquad\xi(v_{k})=\xi(v_{k-1})+\lambda_{k}(v_{k}-v_{k-1}),\qquad\lambda_{k}\neq 0,

where \xi_{0} is the source-coordinate location at the beginning of the window. The local scaling factor \lambda_{k} may vary across steps. The window is valid when 0\leq\xi(v_{k})\leq T for every k. The sign of \lambda_{k} determines the local traversal direction, while |\lambda_{k}| determines the source-coordinate spacing for that step. Constant scaling is the special case \lambda_{k}=\lambda for all k. For positive scaling, a larger value of \lambda_{k} advances farther along the demonstrated trajectory per unit transformed time and therefore corresponds to accelerating along the trajectory; a smaller positive value corresponds to decelerating. When the scale varies across a window, the low-level controller must track these changing effective rates.

Reversal uses the constant schedule \lambda_{k}=-1 for all k, with \xi_{0}=T and \mathcal{V}=T, so that \xi(v)=T-v. Let \mathbf{s}(\xi) denote the full state trajectory. Physical reversal then gives \mathbf{s}^{\mathrm{rev}}(v)=\mathcal{R}\bigl(\mathbf{s}(\xi(v))\bigr)=\mathcal{R}\bigl(\mathbf{s}(T-v)\bigr), where \mathcal{R} preserves positions and negates velocities. For the absolute targets in Section 3.1, reversed demonstrations require actions relabeled toward preceding source configurations, rather than simply reversed action arrays. This provides reverse-task supervision for paired tasks such as opening and closing, but contact, friction, delays, and grasp or release events can make some reversed transitions infeasible.

Positive scaling preserves order. For uniform recording frequency f_{0} and period \Delta t_{0}=1/f_{0}, sampling v every \Delta t_{0} with \lambda=g+1 skips g\in\{0,1,2,\ldots\} source samples. The resulting source-time interval and sampling frequency are

\Delta t_{g}=(g+1)\Delta t_{0},\qquad f_{g}=\frac{f_{0}}{g+1}.

Variable strides give a piecewise time transformation. Thus, reversal changes the task direction represented by supervision, while resampling changes its temporal resolution.

Idealized transition reversibility.

Using the transition model in Section 3.1, write P^{+} and P^{-} for the paired task directions. Following the definition of full time-reversal symmetry in Jiang et al. (2025), an idealized reversibility condition would require that for each transition (s_{t},a_{t},s_{t+1}), there exists a reverse action \hat{a}_{t}\in\mathcal{A} satisfying

P^{+}(s_{t+1}\mid s_{t},a_{t})=P^{-}(\mathcal{R}s_{t}\mid\mathcal{R}s_{t+1},\hat{a}_{t}).

This equality describes an idealized condition for exact transition reversibility; our method does not assume that it holds for every transition in the demonstrations. We nevertheless use reversed trajectories and their relabeled targets directly for training, without segment-level feasibility filtering. This simple augmentation is effective in our experiments (Sections 5.2 and B.2), while filtering infeasible segments is a promising direction for future work.

Strided target prediction.

Building on the DP3 action-chunk formulation in Section 3.3, we now allow the observation history and target samples within a chunk to use positive adjacent offsets rather than the default consecutive spacing. Consider a discrete demonstration indexed by t=0,\ldots,T, where T is the final trajectory index. For an anchor t, we randomly sample observation indices 0\leq i_{1}<\cdots<i_{L}=t and target indices t<j_{1}<\cdots<j_{H}\leq T-1 according to the selected observation and action strides. Consecutive sampling is the special case i_{\ell}=t-L+\ell and j_{k}=t+k. The resulting observation history and action targets are (o_{i_{1}},\ldots,o_{i_{L}}) and (a_{j_{1}},\ldots,a_{j_{H}})=(\psi(o_{j_{1}+1}),\ldots,\psi(o_{j_{H}+1})), respectively, where each action target is the configuration of the next observation.

In the next two subsections, we explain how to exploit time scaling in imitation learning.

4.3 Trajectory Reversal Augmentation

Trajectory reversal augments each task’s expert dataset with demonstrations from its paired task using the offline construction in Section 4.1. The construction reverses the observation order and relabels each action to target the next configuration in the reversed sequence.

Reversed demonstrations.

Given a source trajectory \tau=(o_{0},a_{0},\ldots,a_{T-1},o_{T}) with T transitions, we reverse the observation order:

\tilde{o}_{t}=o_{T-t},\qquad t=0,\ldots,T.

This applies the reversal case \lambda=-1 of Equation 4 to the recorded point clouds and robot configurations.

Using the configuration extractor \psi defined in Section 3.3, the action at each reversed observation targets the next robot configuration in the reversed sequence:

\tilde{a}_{t}=\psi(\tilde{o}_{t+1})=\psi(o_{T-t-1}),\qquad t=0,\ldots,T-1.

The targets are thus obtained directly from recorded proprioception.

Let \operatorname{Rev} denote observation reversal together with this action relabeling. Applying it in both directions yields two augmented datasets, which can be used to train for their respective tasks:

\widetilde{\mathcal{D}}^{+}=\mathcal{D}^{+}\cup\operatorname{Rev}(\mathcal{D}^{-}),\qquad\widetilde{\mathcal{D}}^{-}=\mathcal{D}^{-}\cup\operatorname{Rev}(\mathcal{D}^{+}).

The two datasets reuse the recorded demonstrations without extra environment interaction. We next describe how to sample training windows from these trajectories at different temporal spacings.

4.4 Temporal-Stride Augmentation

For either an original or a reversed trajectory, consecutive training samples correspond to g=0 in Equation 5. Temporal-stride augmentation also samples windows with g\geq 1, so that the policy learns to predict chunks of more widely spaced joint targets. We apply the positive time scaling in Section 4.2 to vary both observation and action spacing while keeping the history length L and action horizon H fixed.

Training windows.

Let \mathcal{B} be the set of allowed positive integer strides. For each training window anchored at index t, we independently and randomly sample the observation and action offset vectors. Let \bm{\delta}^{o}=(\delta^{o}_{1},\ldots,\delta^{o}_{L-1})\in\mathcal{B}^{L-1} and \bm{\delta}^{a}=(\delta^{a}_{1},\ldots,\delta^{a}_{H})\in\mathcal{B}^{H}. These are adjacent offsets: each entry is the gap between neighboring samples, with the first action gap measured from the anchor. They determine the time indices through

i_{L}=t,\kern 5.0pti_{\ell}=i_{\ell+1}-\delta^{o}_{\ell}\kern 5.0pt(\ell=1,\ldots,L-1),\quad j_{0}=t,\kern 5.0ptj_{k}=j_{k-1}+\delta^{a}_{k}\kern 5.0pt(k=1,\ldots,H),

Since all adjacent offsets are positive, these indices satisfy

0\leq i_{1}<\cdots<i_{L}=t=j_{0}<j_{1}<\cdots<j_{H}\leq T-1.

Thus, \bm{\delta}^{o} and \bm{\delta}^{a} directly specify the adjacent temporal spacing of the observation history and action chunk. Writing \eta=[\bm{\delta}^{o},\bm{\delta}^{a}] for this adjacent-offset schedule, we construct

h_{t}^{\eta}=(o_{i_{1}},\ldots,o_{i_{L}}),\qquad A_{t}^{\eta}=(a_{j_{1}},\ldots,a_{j_{H}}).

These offsets describe spacing in the recorded sequence. For uniformly timed recordings, their relation to sampling frequency follows Equation 5.

Temporal-offset conditioning.

Temporal-offset conditioning consists of two offset vectors: the adjacent spacing \bm{\delta}^{o} in the observation history and the adjacent spacing \bm{\delta}^{a} in the future action chunk. We collect them in the schedule \eta=[\bm{\delta}^{o},\bm{\delta}^{a}]. Since these two offset vectors are sampled independently, the observation history alone does not determine the temporal spacing of the action targets. We therefore provide the full schedule \eta to the policy by augmenting the DP3 condition e_{\phi}(h_{t}^{\eta}) with a learned embedding of \eta. Let g_{\omega} denote a learnable function with parameters \omega that encodes the normalized adjacent observation and action offsets. The resulting condition and policy become:

c_{t}^{\eta}=\left[e_{\phi}(h_{t}^{\eta}),g_{\omega}(\eta)\right],\qquad\pi_{\theta}(A_{t}^{\eta}\mid h_{t}^{\eta},\eta),

where brackets denote concatenation and \theta collects the trainable policy parameters.

Inference schedules.

Because the policy is conditioned on temporal offsets, inference must specify the spacing of the future targets before each prediction. At replanning cycle q, the scheduling rule \sigma selects \bm{\delta}^{a}_{q}\in\mathcal{B}^{H} according to a fixed, randomized, or context-dependent strategy. If t_{q} is the current anchor, the target indices satisfy

j_{q,0}=t_{q},\qquad j_{q,k}=j_{q,k-1}+\delta^{a}_{q,k},\qquad k=1,\ldots,H.

For q=0, there is no previous action chunk, so the history is initialized with configured observation offsets \bm{\delta}^{o}_{0}=\bm{\delta}^{o,\mathrm{init}} (consecutive observations correspond to \bm{\delta}^{o,\mathrm{init}}=(1,\ldots,1)), and the initial schedule is \eta_{0}=[\bm{\delta}^{o,\mathrm{init}},\bm{\delta}^{a}_{0}]. For q\geq 1, the current observation offsets are determined by the spacing of the previous action chunk. Although observations arrive one at a time, the retained history represents the gaps induced by the last chunk, so

\bm{\delta}^{o}_{q}=\left(\delta^{a}_{q-1,H-L+2},\ldots,\delta^{a}_{q-1,H}\right).

For example, with L=3, H=6, and \bm{\delta}^{a}_{q-1}=(2,1,3,2,1,3), the transition from cycle q-1 to cycle q gives \bm{\delta}^{o}_{q}=(1,3). The schedule \eta_{q}=[\bm{\delta}^{o}_{q},\bm{\delta}^{a}_{q}] then conditions the prediction via Equation 13. We study the choice of \sigma in the action-offset ablation in Section 5.3.

4.5 RESTA Algorithm

The two augmentations enter training in sequence. We first form \widetilde{\mathcal{D}}^{+} and \widetilde{\mathcal{D}}^{-} using Equation 9. We then sample temporal schedules when constructing windows from these datasets, so both original and reversed demonstrations provide supervision at different spacings.

For direction \kappa\in\{+,-\}, sample a trajectory \tau from \widetilde{\mathcal{D}}^{\kappa}, a valid anchor t, and a schedule \eta. We construct h_{t}^{\eta} and A_{t}^{\eta} using Equation 12 and their condition c_{t}^{\eta} using Equation 13. Using the conditional denoising loss defined in Equation 3, the training objective is

\mathcal{L}(\theta)=\mathbb{E}_{\tau,t,\eta}\left[\mathcal{L}_{\mathrm{DP3}}(\theta;A_{t}^{\eta},c_{t}^{\eta})\right].

This applies the conditional denoising objective in Equation 3 to the augmented windows sampled under the imitation-learning dataset. At inference, the scheduling rule \sigma selects \bm{\delta}^{a}_{q}, which determines the future action indices through Equation 14. We obtain the observation offsets from the preceding chunk using Equation 15. The resulting schedule conditions the next action prediction together with the current observation history. The appendix pseudocode in Algorithm 1 summarizes both stages, and Figure 2 shows the method overview.

5 Experimental Results

We evaluate RESTA on paired bimanual manipulation tasks from the RoboTwin benchmark (Mu et al., 2025; Chen et al., 2025). We first describe the benchmark, training protocol, and evaluation metric in Section 5.1. Section 5.2 compares RESTA with DP3 at the same and twice the demonstration budget. Section 5.3 examines action-offset selection at inference, randomized observation offsets during training, and the individual effects of trajectory reversal and temporal-stride augmentation.

5.1 Experimental Setup

We use eight forward tasks from RoboTwin 2.0 and independently implement one reverse counterpart per task, yielding eight pairs (16 directions; see Section B.1). Each pair has expert datasets \mathcal{D}^{+} and \mathcal{D}^{-}; the former contains forward demonstrations, while the latter is collected in our reverse environment with our expert policy. Offline reversal augments each dataset with the reversed trajectories from its partner, and these synthetic samples do not count toward the data budget. DP3 is the baseline, with fixed observation and action interfaces across methods. All models train for 3,000 epochs and are evaluated for 100 episodes per task; we report per-task and mean success. Settings are listed in Table 3. Unless noted otherwise, temporal-stride/observation-offset methods use the rand schedule, which samples a gap from \{1,2,3\} per action chunk and supplies the corresponding observation offsets.

5.2 Main Results

Evaluation Success Rate

Table 1 compares the baseline and RESTA at two collection budgets. Our goal is to test whether RESTA{}_{n} can improve over the same-budget baseline B_{n} while approaching the performance of the larger-data reference B_{2n}. Across the full task suite, the resulting relationship is B_{n}<RESTA{}_{n}\approx B_{2n}. In practical terms, RESTA can reach the performance of a policy trained with 2n native demonstrations while collecting only n trajectories and deriving additional reversed augmentation trajectories offline. For a paired task, demonstrations from one direction can therefore provide additional supervision for the opposite direction without another collection run. We count only independently collected demonstrations toward the data budget; reversed augmentation trajectories are generated offline.

Execution-Step Analysis

Success rate alone does not reveal the execution cost of successful episodes. We report both policy actions and simulator physics steps until success. Physics steps count each simulator update, including the low-level interpolation used to execute an action. The simulator advances by 1/250 s per physics step, or 4 ms, so dividing the count by 250 gives the simulated duration of an episode. This duration is a closer proxy for robot execution time than the policy-action count, although it does not include policy inference or other system overhead and is not a hardware timing measurement. For each task, we average over successful episodes and then average the task means across the sixteen tasks. In these units, the mean simulated duration is 57.0 s for B_{25} and 37.3 s for \mathrm{RESTA{}}_{25}, and 52.9 s for B_{50} and 36.3 s for \mathrm{RESTA{}}_{50}; B_{100} averages 51.5 s. Thus, RESTA reduces this duration by 19.7 s and 16.6 s at the two matched budgets, respectively. Table 2 reports average policy-action and physics-step counts; per-task results are provided in Tables 8 and 9.

5.3 Ablation Study

We examine how to select action offsets at inference, whether to randomize observation offsets during training, and how each augmentation contributes at different demonstration budgets.

Action-offset Selection

At n=50, we evaluate the same TSAO checkpoint under seven inference schedules. The fixed schedules offset1, offset2, and offset3 use action gaps of one, two, and three, respectively. The rand schedule samples one gap per chunk and applies it to every target, whereas rand_vector samples each gap independently. The voting schedules vote3 and vote11 select among candidate action chunks. This ablation provides the full analysis of action-schedule selection; definitions of all seven inference schedules are provided in Section B.3. At n=50, rand has the highest average success among the seven evaluated schedules (83.2%), so we use it for RESTA and the TSA–TSAO comparison. Full per-task schedule results are in Table 6.

Randomized Observation Offsets

We compare temporal-stride augmentation with consecutive training observations (TSA) against its extension with randomized observation offsets and the corresponding conditioning (TSAO). Both variants randomize action gaps during training and use rand at inference. TSAO exposes the policy to observation histories with different temporal spacings, matching the variation induced by previously executed action chunks. TSAO improves over TSA at n=25 (74.6% vs. 72.4%) and at n=50 (83.2% vs. 80.8%), supporting the benefit of training on varied observation histories and conditioning on their offsets. The full per-task comparison is provided in Table 7.

Effects of the Proposed Techniques

We compare trajectory reversal, temporal-stride augmentation, and their combination at matched collection budgets. \mathrm{TR}_{n} uses n native and n time-reversed trajectories, whereas \mathrm{TSAO}_{n} uses n native trajectories with temporal-stride augmentation and observation-offset conditioning, evaluated with the selected rand schedule. RESTA{}_{n} combines trajectory reversal with this temporal-stride and observation-offset augmentation. The full per-task results and averages are provided in Table 5. \mathrm{TSAO}_{n} is the stronger aggregate component under the randomized inference schedule, consistently improving average success across all evaluated collection settings. \mathrm{TR}_{n} can be more helpful on individual paired tasks whose demonstrations transfer well across directions. Combining the two gives the strongest current-budget result and approaches the larger-data reference without additional independent collection.

6 Conclusion and Limitations

RESTA combines offline paired-task reversal, temporal strides, and temporal offset conditioning. Across 16 simulated RoboTwin tasks, it raises success from 67.5% to 78.4% at 25 demonstrations and from 77.3% to 85.2% at 50, nearing DP3’s 85.4% at 100 while shortening episodes (Tables 1 and 2). Simulation-only tests and infeasible reversals limit our evidence; future work includes automatic task-pair discovery.

AI use statement

We used generative AI tools to assist with drafting and revising selected manuscript text; writing code and running experiments; generating tables from the experimental results; and designing and editing the teaser figure, including its figure-production code. AI tools were not used to fabricate or alter experimental measurements. We reviewed AI-assisted code, experiment outputs, tables, text, and visual materials, and checked reported values against the experiment results. The authors take responsibility for the final content of this work, including text, claims, code, tables, and figures produced with AI assistance.

References

1. Ayyagari et al. (2024) Ranga Shaarad Ayyagari, Anurita Ghosh, and Ambedkar Dukkipati. Temporal abstraction in reinforcement learning with offline data. arXiv preprint arXiv:2407.15241, 2024. URL https://arxiv.org/abs/2407.15241.
2. Bai et al. (2026) Jiaqi Bai, Jindou Jia, Yuxuan Hu, Gen Li, Xiangyu Chen, Tuo An, Kuangji Zuo, and Jianfei Yang. Flash: Efficient visuomotor policy via sparse sampling. arXiv preprint arXiv:2605.15492, 2026. URL https://arxiv.org/abs/2605.15492.
3. Barkley et al. (2023) Brett Barkley, Amy Zhang, and David Fridovich-Keil. An investigation of time reversal symmetry in reinforcement learning. arXiv preprint arXiv:2311.17008, 2023. URL https://arxiv.org/abs/2311.17008.
4. Chen et al. (2025) Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, et al. RoboTwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation. arXiv preprint arXiv:2506.18088, 2025. URL https://arxiv.org/abs/2506.18088.
5. Chi et al. (2023) Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, 2023. URL https://arxiv.org/abs/2303.04137.
6. Ding et al. (2024) Jiayu Ding, Xulin Chen, Garret E. Katz, and Zhenyu Gan. Towards dynamic quadrupedal gaits: A symmetry-guided RL hierarchy enables free gait transitions at varying speeds. arXiv preprint arXiv:2403.10723, 2024. URL https://arxiv.org/abs/2403.10723.
7. Dong et al. (2025) Mingjie Dong, Hanwei Ruan, Zeyu Wang, Chenyang Sun, Shiping Zuo, Yifeng Chen, Jianfeng Li, and Mingming Zhang. Toward physician-level performance in robot-assisted ankle rehabilitation via imitation learning with empirical and temporal adaptation. IEEE Transactions on Robotics, 41:5360–5377, 2025. doi: 10.1109/TRO.2025.3600162.
8. Florensa et al. (2017) Carlos Florensa, David Held, Markus Wulfmeier, Michael Zhang, and Pieter Abbeel. Reverse curriculum generation for reinforcement learning. In Conference on Robot Learning, 2017. URL https://arxiv.org/abs/1707.05300.
9. Hu et al. (2024) Jianshu Hu, Paul Weng, and Yutong Ban. State-novelty guided action persistence in deep reinforcement learning, 2024. URL https://arxiv.org/abs/2409.05433.
10. Jiang et al. (2025) Yunpeng Jiang, Jianshu Hu, Paul Weng, and Yutong Ban. Time reversal symmetry for efficient robotic manipulations in deep reinforcement learning. In Advances in Neural Information Processing Systems 38, pp. 158404–158435. Neural Information Processing Systems Foundation, Inc., 2025. doi: 10.52202/085713-4778. URL https://doi.org/10.52202/085713-4778.
11. Johns (2021) Edward Johns. Coarse-to-fine imitation learning: Robot manipulation from a single demonstration. In 2021 IEEE International Conference on Robotics and Automation (ICRA), 2021. URL https://arxiv.org/abs/2105.06411.
12. Kim et al. (2026) Hyunjun Kim, Jaeseog Won, Hong-ryul Jung, Donghoon Kim, Minsuk Ko, and Hyungpil Moon. Reverse playback motion: A simple way for anyone to obtain high-quality demonstration data. International Journal of Precision Engineering and Manufacturing, 2026. doi: 10.1007/s12541-026-01589-3. URL https://doi.org/10.1007/s12541-026-01589-3.
13. Mu et al. (2025) Yao Mu, Tianxing Chen, Zanxin Chen, Shijia Peng, Zhiqian Lan, Zeyu Gao, Zhixuan Liang, Qiaojun Yu, Yude Zou, Mingkun Xu, Lunkai Lin, Zhiqiang Xie, Mingyu Ding, and Ping Luo. RoboTwin: Dual-arm robot benchmark with generative digital twins. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 27649–27660, 2025.
14. Nair et al. (2020) Suraj Nair, Mohammad Babaeizadeh, Chelsea Finn, Sergey Levine, and Vikash Kumar. Time reversal as self-supervision. In 2018 IEEE International Conference on Robotics and Automation, 2020. doi: 10.1109/ICRA40945.2020.9196862. URL https://arxiv.org/abs/1810.01128.
15. Park et al. (2025) Hyeonjun Park, Daegyu Lim, Seungyeon Kim, and Sumin Park. Proleptic temporal ensemble for improving the speed of robot tasks generated by imitation learning. Journal of Korea Robotics Society, 20(1):61–68, 2025. doi: 10.7746/jkros.2025.20.1.061. URL https://arxiv.org/abs/2410.16981.
16. Qiao et al. (2026) Qiyuan Qiao, Ge Yuan, Can Wang, and Dong Xu. Reverse to advance: Teleoperation-cost effective hard policy learning from reversed easy tasks. arXiv preprint arXiv:2607.13455, 2026. URL https://arxiv.org/abs/2607.13455.
17. Raileanu et al. (2020) Roberta Raileanu, Max Goldstein, Denis Yarats, Ilya Kostrikov, and Rob Fergus. Automatic data augmentation for generalization in deep reinforcement learning. arXiv preprint arXiv:2006.12862, 2020. URL https://arxiv.org/abs/2006.12862.
18. Tagliabue & How (2024) Andrea Tagliabue and Jonathan P. How. Efficient deep learning of robust policies from MPC using imitation and tube-guided data augmentation. IEEE Transactions on Robotics, 40:4301–4321, 2024. doi: 10.1109/TRO.2024.3431988.
19. Tao et al. (2024) Stone Tao, Arth Shukla, Tse-kai Chan, and Hao Su. Reverse forward curriculum learning for extreme sample and demonstration efficiency in reinforcement learning. In International Conference on Learning Representations, 2024. URL https://arxiv.org/abs/2405.03379.
20. Wang & Meger (2021) Haoxu Wang and David Meger. Robotic object manipulation with full-trajectory gan-based imitation learning. In 2021 18th Conference on Robots and Vision (CRV), pp. 57–63, 2021. doi: 10.1109/CRV52889.2021.00016.
21. Wu et al. (2021) Hongmin Wu, Wu Yan, Zhihao Xu, Shuai Li, and Xuefeng Zhou. Learning robot anomaly recovery skills from multiple time-driven demonstrations. Neurocomputing, 464:522–532, 2021. doi: 10.1016/j.neucom.2021.08.036.
22. Yarats et al. (2021) Denis Yarats, Rob Fergus, Alessandro Lazaric, and Lerrel Pinto. Mastering visual continuous control: Improved data-augmented reinforcement learning. arXiv preprint arXiv:2107.09645, 2021. URL https://arxiv.org/abs/2107.09645.
23. Ze et al. (2024) Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. In Proceedings of Robotics: Science and Systems, 2024. URL https://arxiv.org/abs/2403.03954.
24. Zhang et al. (2026) Jiyao Zhang, Zimu Han, Junhan Wang, Xionghao Wu, Shihong Lin, Jinzhou Li, Hongwei Fan, Ruihai Wu, Dongjiang Li, and Hao Dong. Hipolicy: Hierarchical multi-frequency action chunking for policy learning. In Computer Vision – ECCV 2026, 2026. doi: 10.1007/978-3-032-37447-9_20. URL https://arxiv.org/abs/2604.06067.
25. Zhao et al. (2023) Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. In Proceedings of Robotics: Science and Systems, 2023. URL https://arxiv.org/abs/2304.13705.

Appendix A Additional Related Work

Action-Chunk Policies for Imitation Learning

Action-chunk policies predict multiple future targets for receding-horizon execution. Action Chunking with Transformers (ACT) (Zhao et al., 2023) predicts chunks from visual observations, while Diffusion Policy (Chi et al., 2023) models action sequences with conditional denoising diffusion. 3D Diffusion Policy (DP3) (Ze et al., 2024) conditions this formulation on compact point-cloud and proprioceptive features. We use DP3 as our backbone for its geometric and proprioceptive inputs, keeping its architecture fixed to isolate our data augmentations: reversing demonstrations across paired task directions and sampling observation–action windows at varied temporal strides and offsets.

Other methods improve demonstration efficiency or robustness by learning from a single demonstration (Johns, 2021), synthesizing trajectories with GANs (Wang & Meger, 2021), or perturbing MPC-generated data (Tagliabue & How, 2024). DrQ-v2 and DrAC augment visual-control data or select augmentations automatically in reinforcement learning (Yarats et al., 2021; Raileanu et al., 2020). These methods alter observations, state–action coverage, or trajectory content; our method instead changes the temporal indices used to form imitation-learning windows.

Temporal Frequency and Multi-Rate Action Generation

Temporal adaptation has been studied for robot-assisted rehabilitation imitation (Dong et al., 2025), time-driven demonstrations for robot anomaly recovery (Wu et al., 2021), and temporal abstraction in offline RL (Ayyagari et al., 2024). HiPolicy (Zhang et al., 2026) predicts chunks at multiple frequencies, fuses frequency-aligned histories, and selects execution frequency using an entropy-based rule. FLASH (Bai et al., 2026) combines sparse temporal sampling with a continuous Legendre-polynomial action representation for long-horizon flow matching; Proleptic Temporal Ensemble (Park et al., 2025) changes inference-time temporal aggregation to accelerate execution. These methods address different tasks or alter policy representations and inference. Our generic data-level approach changes training-window sampling and supplies the temporal offsets to an existing action-chunk policy, without frequency-specific branches or an entropy-based selector; it improves results under fixed and randomized schedules.

Appendix B Additional Experimental Details

This section reports the DP3 configuration and the temporal-offset settings used in our experiments. All methods use DP3 as the policy backbone; the table separates its model and optimization settings from offsets used by augmented variants.

For completeness, Algorithm 1 gives the combined training and inference procedure.

B.1 Task Suite

The experimental suite contains sixteen bimanual manipulation tasks organized as eight pairs. We select one forward task from RoboTwin 2.0 for each pair and independently implement its reverse-task counterpart. For each pair, we assume expert datasets \mathcal{D}^{+} and \mathcal{D}^{-}; we create the reverse-task environment and implement its expert policy to collect \mathcal{D}^{-}. We then apply trajectory reversal offline in both directions, using \operatorname{Rev}(\mathcal{D}^{-}) to augment \mathcal{D}^{+} and \operatorname{Rev}(\mathcal{D}^{+}) to augment \mathcal{D}^{-}. Thus, trajectory reversal augments the independently collected demonstrations but does not define or replace the reverse task itself. Figure 3 shows expert demonstrations for all sixteen task directions; each sequence contains four frames from the initial state to the final state. The task-pair correspondences are listed in Table 4.

B.2 Effects of the Proposed Techniques

Table 5 reports per-task and average results for the data-efficiency comparison. The effects of the two augmentations vary across tasks. At n=25, trajectory reversal improves Place Empty Cup from 47% to 83%, whereas TSAO reaches 71%. For Open Laptop, TSAO reaches 89%, compared with 81% for trajectory reversal and 70% for the baseline. At n=50, trajectory reversal improves Place Fan from 34% to 58%, but reduces Unhang Mug from 96% to 76%. These task-level differences help explain why reversal yields little aggregate gain at this budget.

B.3 Temporal Schedule Selection

The sweep compares seven schedulers for six-target chunks, with adjacent action gaps from \{1,2,3\}. Their cumulative sums give target distances from the anchor while keeping chunk length fixed. The fixed schedules offset1, offset2, and offset3 repeat gaps of one, two, and three, spanning 6, 12, and 18 steps. rand samples one gap per chunk, whereas rand_vector samples each of the six gaps independently.

The voting schedulers evaluate candidate gap vectors and execute the candidate chunk closest to the mean predicted action chunk. vote3 uses

\mathcal{S}_{3}^{\mathrm{vote}}=\{(1,1,1,1,1,1),(2,2,2,2,2,2),(3,3,3,3,3,3)\},

and vote11 expands this set with eight candidates:

\displaystyle\mathcal{S}_{11}^{\mathrm{vote}}=\mathcal{S}_{3}^{\mathrm{vote}}\cup\{ \\ \displaystyle(1,2,3,1,2,3),(3,2,1,3,2,1), \\ \displaystyle(1,3,2,1,3,2),(2,1,3,2,1,3), \\ \displaystyle(1,1,2,2,3,3),(3,3,2,2,1,1), \\ \displaystyle(1,2,2,3,3,2),(3,2,2,1,1,2)\}.

The eight added vote11 vectors test reordered three-gap cycles, progressive coarsening or refinement, and nonmonotonic spacing. The first six span 12 steps, matching offset2; the final two span 13 and 11 steps. Together with the three constant vectors, they form a compact, non-exhaustive subset of the 3^{6} possible schedules. The executed action gaps set the next cycle’s observation offsets. Comparing vote3 and vote11 tests agreement-based selection, not task success directly; individual vote11 candidates were not ablated.

The preferred fixed gap varies by task: changing it from one to three raises Open Laptop success from 86% to 96%, but lowers Close Microwave from 89% to 71% (Table 6). At n=50, rand has the highest average success (83.2%), versus 81.9% for rand_vector, and is used as the default (Table 6). This favors varying the scale between chunks while keeping regular spacing within each chunk in this evaluation. vote11 reaches 81.9%, up from 81.6% for vote3, but remains below rand.

B.4 Temporal-stride Ablation Details

The action schedule also determines the temporal spacing represented by the next observation history, as specified in Equation 15. This motivates training with varied observation spacings and conditioning on their offsets: the policy must interpret histories produced by different preceding action schedules. Table 7 compares TSA and TSAO using the selected rand schedule at both trajectory budgets. TSAO raises average success from 72.4% to 74.6% at n=25 and from 80.8% to 83.2% at n=50. The improvement is not uniform across tasks or budgets, supporting the combined use of randomized observation spacings and offset conditioning on average; this comparison does not isolate their individual effects.

B.5 Execution-Step Results

The following tables provide per-task execution-step measurements for the baseline and RESTA checkpoints, including policy actions and simulator physics steps. RESTA uses the rand schedule; values are means over ten successful episodes per task.