acceptodds
Withdrawn from ICLR 2027

Not Just Faster: Learning from Mixed-Speed Demonstrations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Demonstrations teach robot policies both task behavior and its timing. Can faster training experience improve task completion without changing deployment settings? Retiming changes how much of a reference trajectory falls within a fixed action window, motivating training on both original and accelerated experience. We investigate this idea through an empirical study of vision-language-action and world-action models across simulated and physical manipulation. Our comparisons distinguish changes in training configuration from changes in a trained policy's speed condition or execution schedule. On CALVIN, mixed-speed configurations improve completion for both model families under the normal condition and unchanged execution settings. Execution sweeps further show that stronger chunk compression can reduce completion and increase control steps among successful trials. These findings support treating demonstration pace as a choice in training supervision, while assessing deployment pace separately through its task-dependent effects on success and execution cost.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Learning from Mixed-Speed Demonstrations

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_WQpVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether faster demonstrations can improve robot learning and whether faster deployment necessarily leads to faster or better task completion. It trains VLA and WAM using a mixture of original speed and retimed demonstrations, and argues that the two speeds provide complementary temporal supervision. Specifically, original speed demonstrations preserve intermediate targets, while accelerated demonstrations expose the policy to longer-horizon progress within the same action window. Experiments on CALVIN, BEHAVIOR-1K, LIBERO, and two real-world insertion tasks show that mixed-speed training generally improves completion, sometimes while reducing the number of control steps or task time. However, execution-time action-chunk compression has task-dependent effects, that is, moderate compression can be beneficial, whereas aggressive compression substantially harms success.

Critical strengths
  1. The paper studies a practically relevant question in robot IL. Its most useful contribution is the clear separation between training pace, behavior conditioning, and deployment pace, which are often conflated in discussions of robot speed.

  2. The evaluation covers multiple model families, simulation benchmarks, and real-world manipulation tasks. Although the results are primarily empirical, the consistent improvement from mixed-speed training across several settings, together with the finding that aggressive execution compression can hurt success, provides useful evidence that training speed and deployment speed should be treated as separate design choices.

Critical weaknesses
  1. The major weakness is that the paper does not cleanly establish that the gains come from mixed-speed supervision itself. The accelerated WAM data are generated through replay and filtered to keep only successful episodes, while the physical mixed-speed policies continue training from original-speed checkpoints. Thus, the improvements could be due to success filtering, broader data coverage, or additional training, rather than the proposed complementary temporal supervision.

  2. The paper’s main explanation is not actually tested. The claim that original- and accelerated-speed demonstrations provide complementary multi-timescale supervision remains largely a hypothesis, since the paper does not measure gradient interactions or otherwise directly validate this proposed mechanism. Combined with the use of a single training seed, success-conditioned step statistics, and approximate real-world timing, the experiments support an interesting empirical observation but not a strong or well-validated technical contribution.

Questions for authors
  1. Can the authors provide a controlled comparison that separates the effect of mixed-speed supervision from success filtering, additional data coverage, and continued training from an original-speed checkpoint?

  2. Do the reported gains persist across multiple training seeds, and are the differences statistically significant given that several key comparisons rely on a single seed?

  3. Is there any direct evidence that the improvements arise from complementary multi-timescale supervision, such as gradient alignment or other mechanism-level analysis, rather than from ordinary data augmentation or exposure to more successful trajectories?

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

The authors propose using mixed-speed demonstrations as complementary learning signals to train generalist robot policies to achieve greater success rates. That is, apart from the original demonstrations collected, they add constructed training data using methods such as TOPP-RA and other hand-engineered ways like interpolation. Using the policies trained with such data, they perform an empirical study and demonstrate high success rates on simulated and real-life robotics tasks under two different foundation model frameworks (VLAs and WAMs). I view the main contribution as essentially a data-augmentation recommendation and an empirical study with the policy gained from using this augmented training data.

Critical strengths

Real-robot evaluations. The authors have demonstrated higher success rates on both simulation and in real-life robotics tasks. Training on mixed-speed demonstrations and doing mixed-speed evaluations could often lead to several mismatches between simulation and reality and it is commendable to have evidence in both scenarios.

Critical weaknesses

Low novelty. The paper doesn’t seem to introduce anything novel, convincingly. The methods used are derived from prior works like TOPP-RA and/or trivially hand-engineered. The claims and insights about using faster demonstrations are already present in prior works like TempoVLA and DemoSpeedup. The proposed data-collection and training strategies are not immediately isolatable as a new system which can be used by practitioners and researchers in this field.

Clarity in both experiment design and presentation. There are several unexplained or underexplained design choices in all the experiment settings used in the paper. I am not able to clearly separate and answer what is done and why it is done in each case. To state concrete examples, for experiments with WAMs, the authors generate faster trajectories with TOPP-RA, and for experiments with VLAs, the authors hand-engineer faster trajectories offline. Also more importantly, in either case, it is never explained as to why this way of collecting new trajectories is essential to the overall story. I also couldn’t find a suitable explanation as to whether actual mixed-speed demonstrations help or if this particular way of demonstration data-augmentation helps. There are also several parameters hand-set, only a singular seed, approximate task time reporting, among others. Since their listed set of contributions are all empirical, I find the haphazard way of experimenting a serious weakness.

Questions for authors

Clarification-based:

  1. Could more details, in clear writing, on the demonstration-augmentation and data collection be provided? And specifically, what was the reason for doing data-augmentation instead of faster demo collection?
  2. While the main contributions are listed in the paper, I would like to know in what way(s) the authors believe practitioners and researchers in this field, can utilize the insights and contributions of this work.

Suggestion-based:

  1. To isolate whether the successes come from data-augmentation and offline trajectory generation that is specific to the work done so far or if it is from mixed-speed demos irrespective of how they are collected, could the authors try an ablation where actual faster demos are collected and mixed into the data instead of doing post-hoc augmentation?
  2. If possible, could the authors perform a smaller scale set of experiments (akin to a sanity check) that are clearly motivated, minimal and well-designed over a singular RL policy training setting (here, the authors may use toy setting likes D4RL or more recent settings like JaxGCRL, or any robotic setting tangible of their choice) and show that mixed-speed demos give considerable success rate improvements?

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

Related papers

Discussion (0)

Sign in to comment.