- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper trains a ~15M-parameter transformer "speed prior" on teleoperated DROID data and handheld FastUMI-100K data. Given a local end-effector path window and the task instruction, it predicts a distribution over execution speed, and a chosen quantile defines a new clock for retiming demonstrations. Retimed DROID episodes are used to fine-tune π0.5, which also predicts "tempo channels" for faster deployment, and the policy is paired with a real-time optimal-control layer (DRAFT/SPLICE/SMOOTH) that respects robot limits and joins action chunks smoothly. On 120 RoboLab tasks the policy beats uniform acceleration at matched speed-up (+3.0 pp success on average, lower jerk and peak speed). On a Franka robot it stacks cups in 3.43 s versus 5.67 s with 30/30 successes.
Critical strengths
-
The idea of learning a task-agnostic speed prior from faster demonstrators (handheld) and transferring it to slower teleoperated data is novel and avoids per-task fitting or annotation. The filtered-BC control (1.04x speed-up despite 1.60x faster demos) convincingly shows that pace transfer, not merely fast data, is what matters.
-
The real-robot results are strong and the system is thoroughly engineered, with a 1.65x speedup at 100% success on cups and a well-specified controller. The ablation showing the PD baseline fails (3% towel success) versus the proposed controller shows that executing human pace requires this component.
Critical weaknesses
-
The real-robot evaluation is narrow and the baselines are not very strong: three tasks, 5 episodes for trash cleaning, and no comparison against the closest prior methods (SAIL, DemoSpeedup, RACE, Realtime-VLA V2) run on the same setup. The SAIL comparison is a cross-paper number on a different robot, so the "4.3x faster" claim is not controlled, and the 30-trial success rates (e.g., 0.90 vs 0.97) are too noisy to distinguish conditions.
-
The contribution bundles a speed prior, tempo channels, and a new controller, and the ablations do not cleanly separate their effects; the default policy under the proposed controller already gets most of the gains in several cases. In RoboLab the controller doesn’t seem to be used at all, and the success gain is modest, while the "half the jerk" claim holds at one matched point (2.8x) and partly reflects the retimed demonstrations being smoother, as the paper itself notes.
Questions for authors
- Can you compare directly against at least one prior speedup method (e.g., SAIL or RACE) on the same robot and tasks, and provide statistics for the differences at 30 trials?
- How much of the real-robot gain comes from the controller versus the retimed data? Please run the default and expert policies under the PD baseline and under your controller across all tasks, and run Ours under PD.
- Is there any evidence of failure modes where the transferred pace is unsafe, such as on tasks with contact-rich or precision phases?
- How sensitive are the results to the choice of target domain (shortest quarter, FastUMI) and to the DCT low-pass filter, given the tracking-jitter confound?
3 comments
Q2. Controller versus retimed data (W2, Q2)
Thank you for suggesting this decomposition. On the two tasks above, each policy now runs under both controllers. The table reports UPH.
| Policy | Bananas, PD | Bananas, our controller | Cubes, PD | Cubes, our controller |
|---|---|---|---|---|
| Default (1x) | 200.7 | 231.7 | 234.0 | 354.0 |
| Ours (1x) | 266.5 | 376.7 | 284.0 | 446.7 |
| Ours (q99) | 394.0 | 492.9 | 338.0 | 511.1 |
Each column shows the effect of the retimed data under one controller, and each row the effect of the controller for one policy. Both raise UPH, and neither alone reaches what the two give together.
As you suggested, we also ran every policy under PD on the fine-tuning tasks (successful episodes out of 30).
| Policy | Cups, PD | Cups, our controller | Towel, PD | Towel, our controller |
|---|---|---|---|---|
| Default (1x) | 28 | 29 | 4 | 29 |
| Expert (1x) | 9 | 9 | 0 | 18 |
| Ours (1x) | 23 | 29 | 0 | 29 |
| Ours (q99) | 11 | 28 | 0 | 30 |
Our policies predict the positions that the arm should reach, and a faster policy places these positions further apart. The PD baseline cannot follow them at this pace. The arm falls up to 20° behind, against about 1° with our controller, and the task fails. On cleaning trash, we stopped the runs under PD for safety.
RoboLab does not use our controller, so the gain in success rate over uniform acceleration at the same speed-up comes from the retimed data alone (95% CIs, Figure 4).
| Ours | q50 (1.35x) | q90 (1.81x) | q95 (2.00x) | q99 (2.38x) |
|---|---|---|---|---|
| Gain in SR | +0.035 [0.003, 0.067] | +0.039 [0.011, 0.068] | +0.044 [0.018, 0.071] | +0.078 [0.048, 0.110] |
The revised Table 1 compares the motion at two matched speed-ups, 2.0x and 2.4x (jerk 0.83 against 1.23 and 1.16 against 2.41). As you note, this smoothness comes from the retimed demonstrations, and this is intended. Retiming with the median clock raises the SPARC of held-out DROID episodes from −3.91 to −3.09, whereas uniform acceleration to the same duration leaves it at −3.87 (Section 5.5).
Q3. Failure modes where the transferred pace is unsafe (Q3)
We counted the events of the contact and precision phases in all 1,200 episodes of each policy on the 120 RoboLab tasks (Appendix E, Table 7).
| Default (1x) | Default (2.5x) | Ours (q99) | |
|---|---|---|---|
| Grasps of a target lost within 1 s | 54.8% | 68.7% | 64.5% |
| Episodes with a target knocked out of the scene | 2.3% | 4.4% | 2.7% |
| Gripper-table collisions per episode | 0.99 | 1.98 | 1.94 |
| Bumps into objects per episode | 3.91 | 5.05 | 4.65 |
A faster pace has a cost for both policies, but it is not larger for the transferred pace than for uniform acceleration at a similar speed-up, and a target object is knocked out of the scene less often.
On the real robot, our policy did not trigger a safety reflex of the arm in any of its 240 zero-shot episodes. On stacking cups, which requires precise placement, its success rate stays between 0.93 and 0.97 on all clocks, against 0.97 for the default policy (Table 3).
Q4. Sensitivity to the target domain and to the DCT (Q4)
When we retime DROID, the target domain could be DROID itself, a fast subset of DROID, or another dataset from faster demonstrators. With DROID itself, the prior reaches a speed-up of only 1.60x at its 0.99 quantile, against 2.38x for ours, and its policy is 0.032 below ours in success rate at the same speed-up (95% CI [0.002, 0.063], Table 14). A fast subset of DROID would be selected by task as much as by demonstrator, because its episodes differ in task. We therefore use another dataset, FastUMI-100K. When we fine-tune on a single task, all episodes share the task, so we use a fast subset, the shortest quarter. It was fixed before any evaluation, and we did not train policies on other fractions.
The setting of the DCT was fixed from the noise level of the tracked FastUMI-100K poses and was not tuned on results.
| Prior | Speed-up (median clock) | SPARC (retimed episodes) | Speed-up with tracking jitter |
|---|---|---|---|
| With DCT | 1.28x | −3.09 | 1.48x |
| Without DCT | 1.56x | −3.45 | 2.04x |
Without the DCT, the prior is faster but less smooth and less robust to tracking jitter. With it, the residual of the prior is uncorrelated with the removed energy (Spearman correlation 0.008 on DROID and 0.001 on FastUMI-100K), so the transferred pace mainly reflects how handheld demonstrators move rather than how they are tracked (Appendix H).
Once again, we really appreciate your insightful questions. We hope that our response addresses all of your questions.
References
[1] Nadun Ranawaka Arachchige et al., SAIL: Faster-than-Demonstration Execution of Imitation Learning Policies, CoRL 2025
[2] Sunwoo Kim et al., Time Optimal Execution of Action Chunk Policies Beyond Demonstration Speed, ICLR 2026
Gentle Reminder
Dear Reviewer TuTf,
As the discussion period ends soon, we wanted to kindly check whether our response has addressed your concerns. We would be happy to clarify anything further.
Best regards,
The Authors
Thank you for your constructive feedback. We appreciate your positive remarks on the novelty of learning a task-agnostic speed prior from faster demonstrators, the filtered-BC control, and the engineering of the system. Our responses to your specific questions are provided below. The numbers are from the revised pipeline described in our global response.
Q1. Comparison with prior speed-up methods on the same robot and tasks (W1, Q1)
We agree that the cross-paper comparison with SAIL was not controlled, and we have removed it from the main text. We implemented SAIL [1] and RACE [2] on the same policy backbone and data as the default policy and ran them on our robot, a Franka Research 3 with a Robotiq 2F-85 gripper. The two tasks are placing two bananas into a tray within 30 s and nine cubes into a bin within 60 s, without any fine-tuning on them. The cube task is new. Every condition runs 30 episodes (Appendix F, Tables 8 and 9).
On its 0.99 tempo channel, our policy places more items than both methods on both tasks. We agree that success rates from 30 trials do not separate conditions, so we test the number of items placed per episode with a two-sided Mann-Whitney U test. The difference is significant against SAIL (p = 0.011 on bananas and 0.012 on cubes) and against RACE (p = 0.011 and p < 0.001). SAIL is reported at 3x because at 5x it reaches a lower UPH on bananas (272.9 against 303.8) and could not be run safely on cubes.
We also ran both methods on the three fine-tuning tasks. The table reports the successful episodes out of 30 and their mean completion time on stacking cups and folding towel, and the items placed per hour on cleaning trash (Appendix G, Table 10).
SAIL succeeds less often than the default policy on stacking cups and folding towel. Our policy on its 0.99 tempo channel succeeds as often as the default policy, in less time than SAIL, and places more trash per hour. Counting a failed episode as the slowest, its completion time over all 30 episodes is shorter than that of the default policy and of SAIL on both tasks (p < 0.001 in the same test). On folding towel and cleaning trash, the motion of RACE was unsafe for the robot, and we could not complete the runs.
In simulation (RoboLab, 120 tasks), the comparison is the following, where TPR is the throughput with regret per minute, which rewards a fast success and penalizes a failure (Appendix E, Table 6).
Raising the setting of SAIL does not take it beyond a speed-up of 1.81x, where it is 0.053 below ours in success rate (95% CI [0.021, 0.083]) and 0.75 below in TPR (95% CI [0.32, 1.15]). RACE falls below the default policy at 1x in success rate.
DemoSpeedup, like SAIL, finds the segments that need precision and plays the others faster by a fixed factor, and it sets this factor by hand for each task. We ran SAIL as the method of this kind. Realtime-VLA V2 trains a speed model on the adjustments that an operator makes by hand to a throttle while the policy runs on the robot. It needs new rollouts with an operator, which do not exist for tasks that the policy has not seen, so we did not run it.