acceptodds
Withdrawn from ICLR 2027

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

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

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 **Re**versal and **S**tride-based **T**rajectory **A**ugmentation (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.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robotics control, imitation learning, time scaling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DXmnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript introduces a time-scaling data augmentation framework called RESTA to improve data efficiency in robot imitation learning. RESTA combines two techniques: trajectory reversal, which reuses demonstrations across paired tasks with opposite objectives, and temporal-stride augmentation, which varies observation and action sampling intervals while conditioning the policy on temporal offsets. Built upon DP3, RESTA requires no additional environment interaction. Experiments on 16 paired bimanual manipulation tasks demonstrate improved success rates and reduced simulated execution time. With 50 demonstrations per task, RESTA achieves an 85.2% average success rate, comparable to the 85.4% achieved by the baseline using 100 demonstrations. Ablation studies show that temporal-stride augmentation provides more consistent improvements, while trajectory reversal is particularly beneficial in low-data settings but may introduce negative transfer in certain tasks. The main limitation is that reversed trajectories are not always physically feasible, especially in contact-rich manipulation.

Critical strengths

RESTA adds data-level transformations and temporal-offset conditioning to an existing action-chunk policy. The paper compares reversal-only, temporal-stride variants and the full method, and evaluates multiple inference offset schedules (Tables 5-7). This helps identify where the improvements originate, even though the effects are not uniform across tasks. The improvements over DP3 are substantial across the 16-task suite (+10.9 and +7.9 percentage points at 25 and 50 demonstrations, respectively). The reported results indicate a practical benefit from reusing existing demonstrations instead of collecting more task-specific data (Table 1).

Critical weaknesses

The reported training configuration specifies a single seed (seed 0), and the paper does not provide training-seed variability or confidence intervals. Comparisons rely primarily on the DP3 baseline and related ablations, with no controlled comparison to the cited alternative temporal-frequency methods. Moreover, evaluation is simulation-only. These limitations make it difficult to assess whether smaller differences, such as 85.2% versus 85.4%, are meaningful or whether the benefits transfer beyond this setting (Section 5; Tables 1, 3 and 5).

Questions for authors

Could the authors provide results across multiple training seeds, including standard deviations or confidence intervals? This would help establish whether the reported improvements, particularly the smaller differences between RESTA and TSAO, are statistically reliable. How does RESTA compare with related temporal-frequency approaches, such as HiPolicy and FLASH, under comparable experimental settings? A controlled comparison, or a clear explanation of why such comparisons are infeasible, would help assess the method's relative advantages.

Reviewer_2GXoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes RESTA, a method for improving demonstration efficiency of robot imitation learning. It applies action-chunk imitation policies that squeeze more supervision out of each recorded demonstration by manipulating time. RESTA combines two transformations. (1) Paired-task trajectory reversal, which reverses observations from an opposite-direction task and relabels actions using recorded robot configurations. and (2) Temporal-stride augmentation, which independently samples adjacent observation and action gaps from {1, 2, 3} for each training window and conditions the policy on these temporal offsets. The method is evaluated on 8 paired bimanual manipulation problems, comprising 16 task directions in simulation. With 25 and 50 demonstrations per task, RESTA achieves average success rates of 78.4% and 85.2%, respectively, compared with 67.5% and 77.3% for DP3, a 11% improvements. The paper also showed a shorter simulation execution durations and ablations of reversal.

Critical strengths
  • The authors use a simple, low-cost, backbone-agnostics way to reuse the temporal structure in demonstrations. Therefore, RESTA reuses existing paired-task demonstrations without additional data collection, environment interaction, or a learned dynamic model, only requires a lightweight temporal-offset embedding. It improves the average success over DP3 by 10.9 with 25 demonstrations and 7.9 with 50 demonstrations. Its simplicity and reported gains make it a potential useful contribution for practitioners problems.

  • Temporal-offset conditioning provides a coherent way to learn from demonstration at different resolutions. RESTA varies the spacing of observation histories and future action targets while explicitly informing the policy of the offsets, which allows the policy to interpret and predict at different temporal spacings.

Critical weaknesses
  • The ablation show that reversal helps some task while hurts others. Table 5 separates reversal from temporal-stride augmentation and reports substantial task-dependent differences. Reversal mainly helps the harder forward tasks, while offering little benefit or causing negative transfer in the reverse direction. The paper does not explain this difference. Different baseline success rates and unclear demonstration sources make it difficult to determine when reversal is useful.

  • The experiments may need stronger controls to establish the gains. All models are running on the same training seed, so additional runs are needed to assess how consistently the improvements hold. Some differences are small, RESTA50 improves over TSAO50 by only 2%, while reversal alone provides almost no improvement at that budget. More training seeds, matched training updates, and a simple temporal-subsampling baseline could make the conclusion more convincing.

Questions for authors
  • Reverse-task construction. How were the eight reverse-task environments and their native demonstrations created? Were the demonstrations collected independently by a scripted expert, or derived from forward demonstrations? How were the reverse tasks' initial state distribution and success conditions checked?

  • Reliability of gains. Could you report the main baseline, TSAO, and RESTA comparisons across multiple training seeds, with the number of gradient updates matched, to show whether the improvements are consistent?

Reviewer_btW1OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RESTA, a data-augmentation framework for improving the data efficiency of robot imitation learning through trajectory reversal and temporal resampling. For paired tasks with opposite objectives, the method reverses observation sequences and relabels actions using recorded robot configurations to construct additional supervision. It also varies the sampling intervals of observation histories and future action targets. Experiments showed that RESTA improves average success rates.

Critical strengths
  • Trajectory reversal and temporal resampling require no additional environment interaction.

  • The paper explicitly treats reversal as an approximate paired-task prior rather than an exact physical symmetry. It recognizes that contact, friction, and grasping or release events can make reversed transitions infeasible, providing a more careful justification for the augmentation.

Critical weaknesses
  • The method primarily combines trajectory reversal, temporal resampling, and offset conditioning. Although this combination is practical, the paper does not sufficiently establish its technical advantage beyond existing augmentation and temporal-sampling approaches. Further, the training protocol does not establish equal optimization budgets. Fixing training to 3,000 epochs does not necessarily provide a fair comparison when reversal increases the number of training trajectories. If an epoch traverses the augmented dataset, RESTA may receive more gradient updates than the baseline. If each epoch instead uses a fixed number of sampled windows, this should be stated. Reporting optimizer steps, sample exposure, and training time, together with continued-training or repeated-sampling baselines, is necessary to separate augmentation benefits from additional optimization

  • All experiments are conducted in simulation, leaving the effectiveness of RESTA on physical robots unverified. This is particularly relevant because trajectory reversal and temporal resampling may produce targets whose feasibility depends on contact dynamics, friction, actuator limits, and controller tracking behavior.

  • Equation (9) combines native demonstrations with reversed demonstrations from the opposite task direction, while the experimental description also refers to constructing reverse-task counterparts through reversal. The paper should explicitly distinguish independently collected trajectories from their derived versions and report the total collection budget per task pair. It should also clarify whether dataset splits are constructed before augmentation, with all versions of an original trajectory kept together.

Questions for authors

see weakness

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.