acceptodds
Under review as a conference paper at ICLR 2027

FreeSpeed: Training-Free Online Speed Modulation for Action-Chunking Policies

Abstract

Online control of execution speed is essential for deploying robot policies in real-world scenarios, as robots may need to speed up under time constraints or slow down to facilitate human interaction and improve safety. However, imitation-learned policies inherit the execution speed of their demonstrations, and test-time speed modification can introduce unrecoverable out-of-distribution observations, reducing task success. We introduce FreeSpeed, a training-free module that post-processes action chunks from pretrained policies. FreeSpeed resamples each predicted chunk at the requested rate, then uses directional inconsistency between adjacent actions as the primary signal for rescaling. This signal adaptively determines how closely the execution speed can approach the requested speed, allowing flexible speed adjustment within the evaluated limits without compromising task success. Across three policy families and 50 simulated tasks, FreeSpeed supports online speed changes, with realized execution rates spanning 0.22x to 2.53x among settings that preserve per-task success. Across four real-world manipulation tasks, FreeSpeed achieves an average success rate of 94%, matching the frozen policy's 93.8%, while realizing execution rates from 0.38x to 1.97x.

Primary area:
applications to robotics, autonomy, planning
Keywords:
execution speed modulation, training-free test-time control, imitation learning, action-chunking policies

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_B9W9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a training-free approach (FreeSpeed) for modulating the robot execution speed. The paper is motivated by the fact that naively resampling the predicted action chunk may lead to out-of-distribution states and affect the policy performance. Therefore, FreeSpeed uses the inconsistency between adjacent actions within a single chunk as the primary signal for action chunk transformation. Experiments under both simulation and real-world experiments demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The paper's perspective is interesting, the method is simple, and the experiment results demonstrate the effectiveness of the proposed method.

  2. The paper is well-written, with ideas and experiment results demonstrated clearly.

Critical weaknesses

While I lean towards acceptance of the paper, I have the following concerns:

  1. I am confused why FreeSpeed can sometimes improve the performance over the reference model when accelerated or decelerated. For deceleration, I suspect it might be due to that the upsampling process does some smoothing and can somehow improve. But acceleration performs downsampling, and rescaling is actually introducing new actions that may deviate from the training distribution, so that the robot state may have a higher probability of falling into an OOD scenario. Moreover, Figure 4 indicates that scaling mostly happens at free-motion stages (instead of physical interaction stages). These stages, under pick and place tasks, should just be using the robot to approach the object or carrying the object around. Thus, I do not see how these stages could affect performance, and would like the authors to clarify more on these aspects.

  2. FreeSpeed seems to be sensitive to be very sensitive to the choice of decay rate λ\lambda. The generalizability of the method remains questionable, as a good range of λ\lambda could vary substantially over different benchmarks, requiring much parameter tuning.

Questions for authors
  1. I am not sure about the practicability of 'slowing down' for robots. Most robots already execute very slowly (also as this paper's real-world demonstrations show) , slowing them done further does not seem necessary. The upsampling process seems more like a smoothing technique. Also, the paper does not clearly explain what are the upsampling and downsampling algorithms they use.

  2. For Figure 4(b), the acceleration case could degenerate to the 1x reference prediction, while the deceleration one cannot. What is the purpose of design that the deceleration one cannot converge to the factor of 1?

8 comments
AuthorsOpenReview ↗

Response to reviewer B9W9 (1/6)

We sincerely thank Reviewer B9W9 for the insightful and constructive comments. We are glad that you recognize the perspective of this work, the simplicity of the method, and the validation of its effectiveness in simulation and real-robot experiments, and we appreciate your positive assessment of the writing. Below we respond to each weakness and question, and we hope that these responses address your concerns. A revised manuscript has also been uploaded, with all additions and changes marked in blue.

Weakness 1: I am confused why FreeSpeed can sometimes improve the performance over the reference model when accelerated or decelerated. For deceleration, I suspect it might be due to that the upsampling process does some smoothing and can somehow improve. But acceleration performs downsampling, and rescaling is actually introducing new actions that may deviate from the training distribution, so that the robot state may have a higher probability of falling into an OOD scenario. Moreover, Figure 4 indicates that scaling mostly happens at free-motion stages (instead of physical interaction stages). These stages, under pick and place tasks, should just be using the robot to approach the object or carrying the object around. Thus, I do not see how these stages could affect performance, and would like the authors to clarify more on these aspects.

Response: Thank you for your careful reading and comments. Your questions help us clarify the role of rescaling and how free motion affects task success. Your comment raises three points, and we respond to each in turn, quoting the corresponding sentences before each response. They concern whether rescaling pushes the robot out of distribution, why FreeSpeed sometimes exceeds the reference, and how the free-motion phases can affect performance. The paired analysis behind the second and third points has been added to Appendix E.7 of the revised manuscript.

In response to: "But acceleration performs downsampling, and rescaling is actually introducing new actions that may deviate from the training distribution, so that the robot state may have a higher probability of falling into an OOD scenario. Moreover, Figure 4 indicates that scaling mostly happens at free-motion stages (instead of physical interaction stages)."

1) FreeSpeed limits the enlargement of the step length under acceleration.

Please allow us to first clarify the acceleration logic of FreeSpeed. Under acceleration, FreeSpeed downsamples the action chunk predicted by the policy. Under the 2× command, for example, the number of steps executed per chunk is roughly halved. To cover the displacement of the original chunk in fewer control steps, the skipped motion has to be compensated. Otherwise, downsampling alone would only discard part of the trajectory. The policy would then need two inferences to complete the original motion, which takes the same number of action steps and, at the same control frequency, the same execution time. For this reason, both baselines by default compensate the full displacement, which is equivalent to a rescaling factor of 2 under the 2× command. Vanilla resampling adds the skipped increments to the retained ones, and w/o CR multiplies the retained increments by the commanded rate.

FreeSpeed uses the same downsampling stride but limits the rescaling factor to between 1 and 1.8 according to cosine inconsistency. When the directions are consistent, the factor is close to 1.8. As the inconsistency increases, the factor moves toward 1. Relative to the factor of 2 used by w/o CR, FreeSpeed therefore limits the enlargement of the step length rather than increasing it from the original 1× step length. Any test-time change of speed moves the policy away from the 1× reference, so every such method is out of distribution to some degree, and FreeSpeed is designed to keep this shift as small as possible. Our analysis in Section 5.5 (Figure 6 and Appendix E.4) shows that keeping the step length close to the original prediction keeps the observations close to the training distribution.

Your observation about Figure 4 is accurate. Relative to the original increments predicted by the policy, the step length is enlarged mainly in the free-motion phases. In our hypothesis, these phases are exactly where acceleration is safe and does not lower the success rate, that is, where test-time modulation yields a free gain in speed. Taking the nominal displacement compensation of a factor of 2 as the reference, FreeSpeed lowers the factor to close to 1 near grasping and placement, so the suppression of step-length enlargement is stronger in these phases.

AuthorsOpenReview ↗

Response to reviewer B9W9 (2/6)

Response to Weakness 1 (continued):

The observation-distribution analysis in Figure 6b supports the effect of this modulation. On Task 60 under the 2× command, the median distance from the observations of FreeSpeed to the training data is 1.09 times that of the reference, compared with 1.79 times for vanilla resampling and 1.68 times for w/o CR. In this analysis, the observation distribution shift of FreeSpeed is therefore smaller.

In response to: "I am confused why FreeSpeed can sometimes improve the performance over the reference model when accelerated or decelerated. For deceleration, I suspect it might be due to that the upsampling process does some smoothing and can somehow improve."

2) Higher success on some tasks is accompanied by lower success on others.

You observe that FreeSpeed achieves a higher success rate than the reference on some tasks. Table 2 also contains tasks on which it is lower than the reference, and the mean success rate over the ten tasks is close to the reference.

The paired outcomes from identical initial states further illustrate this. Over the 500 initial states, the 1× reference succeeds on 444 and FreeSpeed under the 2× command on 443, so the two success rates are statistically the same. There are 36 initial states on which the reference fails and FreeSpeed succeeds, and 37 on which the reference succeeds and FreeSpeed fails. The sets of initial states on which the two methods succeed therefore differ, but the changes in the two directions largely cancel overall. The same holds under deceleration. FreeSpeed performs no interpolation or smoothing under deceleration. It only splits each predicted increment into smaller steps along its original direction, so results above the reference on individual tasks also belong to this two-way change. The cause of this change is explained in point 3.

In response to: "These stages, under pick and place tasks, should just be using the robot to approach the object or carrying the object around. Thus, I do not see how these stages could affect performance, and would like the authors to clarify more on these aspects."

3) The free-motion phase affects the state at the subsequent grasp.

As described above, the modification that FreeSpeed applies during free motion does not directly cause an out-of-distribution state that would make the task fail. If your question is why modifying the speed in this phase can still lead to a higher success rate than the reference, our analysis of individual cases suggests the following explanation.

A change in execution during the approach phase changes the robot state and the observation at each replan, which in turn affects the actions predicted by the policy and the position at which the gripper closes. Even if the rescaling factor is close to 1 near the grasp, the state in which the robot enters this phase may already differ from the reference. The gripper command is not rescaled in magnitude, but this does not guarantee that the gripper closes at the same time and position.

Task 69 shows the largest increase in success among the ten tasks, from 80% for the reference to 94% for FreeSpeed. FreeSpeed succeeds from nine initial states on which the reference fails, and it fails from two on which the reference succeeds. Figure A8 in Appendix E.7 shows both directions of outcome change through two pairs of keyframes on this task. From initial state 20, the reference closes the gripper beside the box, with its fingers closing on empty space, whereas FreeSpeed grasps the box. From initial state 32, the box slips out after FreeSpeed closes off-center, whereas the reference succeeds.

These examples offer one possible explanation for the two-way change in the paired outcomes. A change in the approach can align a grasp that would otherwise miss the object, and it can also make a grasp fail that would otherwise succeed. Free motion therefore does not complete the grasp directly, but it can still affect task success by changing the state at the subsequent interaction.

We would be glad to clarify further if you have any other questions.

Revision: Appendix E.7 adds the paired analysis of success and failure from identical initial states, and Figure A8 shows the paired keyframes of Task 69 from initial states 20 and 32.

AuthorsOpenReview ↗

Response to reviewer B9W9 (3/6)

Weakness 2: FreeSpeed seems to be sensitive to be very sensitive to the choice of decay rate λ\lambda. The generalizability of the method remains questionable, as a good range of λ\lambda could vary substantially over different benchmarks, requiring much parameter tuning.

Response: Thank you for your careful reading and comments. Regarding the sensitivity and generalizability of the decay rate λ\lambda, we make two points. First, all evaluations in this paper use the same fixed λ\lambda, which is not adjusted per benchmark or policy. Second, the single-task sweep in Figure 7b shows a wide effective range of λ\lambda, within which both the success rate and the actual acceleration are maintained. The revised manuscript adds to Section 5.7 the task used in the sweep, the result of the default value on this task, and a statement that all constants are kept the same across all experiments.

1) The fixed parameters are used directly in all evaluation settings. λ=50\lambda = 50 was determined on three LIBERO-90 tasks, Tasks 2, 18, and 34. All results reported afterwards across benchmarks, policies, and the real robot were obtained with this fixed setting, without readjusting λ\lambda or the endpoint mappings for any evaluation environment. Specifically, the remaining LIBERO-90 tasks with Flow Matching, the four LIBERO suites with π0.5\pi_{0.5} and Fast-WAM, Robomimic with Diffusion Policy, the four real-robot tasks, and the seven commanded rates all use the same set of constants. The transfer of the method across these settings therefore does not rely on per-benchmark tuning.

2) The single-task sweep shows a wide effective range.

Figure 7b sweeps λ\lambda on LIBERO-90 Task 2 under the 2× command, with 50 rollouts per value and a logarithmic horizontal axis. The default setting λ=50\lambda = 50 gives a success rate of 88% and a realized rate of 1.28×, matching the result in Table 2. As λ\lambda varies from 22 to 140, the success rate ranges from 86% to 98% and the realized rate from 1.18× to 1.31×. On this task, the effective values are therefore not confined to a narrow interval around the default, and actual acceleration is retained within this range.

We would be glad to clarify further if you have any other questions.

Revision: The last paragraph of Section 5.7 now states the task used in the sensitivity analysis and its results, and that all constants are identical across all experiments.

Reviewer_VQ2fOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces FreeSpeed, a training-free module for controlling the execution speed of frozen action-chunking policies. It combines temporal resampling with adaptive action rescaling, using cosine inconsistency between adjacent translation increments to determine how aggressively to modify step lengths. The method supports acceleration, deceleration, and changes to user-specified rates during execution. Evaluations across three policy families, 50 simulated tasks, and four real-robot tasks demonstrate practical speed modulation while largely preserving task success.

Critical strengths
  1. A practical contribution with a coherent motivation. The setting is closely related to SuP, but FreeSpeed eliminates the need to train an additional scheduling policy. Its support for user-specified rates and online rate changes also provides a more flexible control interface.

  2. An intuitive mechanism with some novelty. Cosine inconsistency provides a simple, learning-free signal for adapting action rescaling. The intuition that directionally consistent motion tolerates more aggressive speed modification is reasonable, and the observation-distribution analysis provides supporting empirical evidence.

  3. A reasonably comprehensive evaluation. The experiments cover multiple policy families, simulation and real-world execution, acceleration and deceleration, and changing rate commands. The component ablations and sensitivity analyses of (fsf_s) and (λ\lambda) help substantiate the design choices and characterize the method’s current capabilities.

Critical weaknesses
  1. Limited evidence of acceleration during precision-demanding operations. Although the evaluation includes grasping, placement, and stacking, it does not clearly establish whether the precision-critical phases themselves are accelerated. The reported task-level speedups may primarily come from less constrained transit motion, while conservative rescaling preserves success near critical interactions.

  2. Limited expansion of demonstrated task capabilities. The method offers a useful simplification and greater flexibility, but the experiments do not establish acceleration capabilities beyond the task regimes already explored by prior work. Evidence on tighter-tolerance manipulation would strengthen the empirical contribution.

Questions for authors
  1. Could the authors provide phase-wise execution times and realized speedups, separating transit motion from approach, grasping, placement, and release? This would clarify where the acceleration gains originate.

  2. How well does cosine inconsistency capture criticality when precise execution requires directionally consistent motion, such as tight-tolerance insertion? Evaluation or failure analysis on such tasks would help establish the limits of the proposed signal.

5 comments
AuthorsOpenReview ↗

Response to reviewer VQ2f (1/3)

We sincerely thank you for your insightful and constructive comments. We are glad that you recognize the practical contribution of this work, the use of cosine inconsistency as a learning-free signal, and the breadth of the evaluation, which covers multiple policy families, simulation and the real robot, and both acceleration and deceleration. Below we respond to each weakness and question, and we hope that these responses address your concerns. A revised manuscript has also been uploaded, with all additions and changes marked in blue.

Weakness 1: Limited evidence of acceleration during precision-demanding operations. Although the evaluation includes grasping, placement, and stacking, it does not clearly establish whether the precision-critical phases themselves are accelerated. The reported task-level speedups may primarily come from less constrained transit motion, while conservative rescaling preserves success near critical interactions.

Question 1: Could the authors provide phase-wise execution times and realized speedups, separating transit motion from approach, grasping, placement, and release? This would clarify where the acceleration gains originate.

Response: Thank you for this insightful comment. Weakness 1 and Question 1 concern the same point, so we address them together. The revised manuscript adds Appendix E.5, which reports the realized rate of each phase (Table A15) and shows how cosine inconsistency and the rescaling factor vary across task phases (Figure A5).

Your assessment is accurate. This is in fact the intended behavior. We found that accelerating in the precision-critical phases can directly cause task failure, so we intentionally place the acceleration in the transit motions, that is, where cosine inconsistency is small. As a result, FreeSpeed gains most of its acceleration from the less constrained approach and transport motions, while rescaling becomes conservative near critical interactions. The table below gives the realized rate of each phase on LIBERO-90 Task 18 (put the frying pan on the stove) under the 2× command, computed over the 48 initial states on which both the reference and FreeSpeed succeed. Grasping and placing are processes rather than single gripper events, so we delimit them by cosine inconsistency. The grasp phase is the period around the gripper closing in which cosine inconsistency is at least 0.05, and it lasts 0.9 s on average. The placement phase is the period at the end of the rollout in which cosine inconsistency is at least 0.05. It lasts 0.4 s on average and appears in 29 of the 48 rollouts. The reference rollout from the same initial state is split at the same positions. The phase-wise presentation follows SuP [1].

PhaseRealized rate
Approach1.35×
Grasp0.84×
Transport1.54×
Placement1.13×
Full rollout1.34×

The approach and the transport realize 1.35× and 1.54×, whereas the grasp and the placement realize 0.84× and 1.13×. Figure A5 shows how cosine inconsistency and the rescaling factor vary across the phases of the same task. Cosine inconsistency peaks near the grasp and rises again at the placement, where the rescaling factor is close to 1.

Revision: Appendix E.5 now includes Table A15, which reports the realized rate of each phase on Task 18 under the 2× command, and Figure A5, which shows how cosine inconsistency and the rescaling factor vary across the phases of the same task.

References:

[1] Z. Wu, et al. "Speedup Patch: Learning a plug-and-play policy to accelerate embodied manipulation." International Conference on Machine Learning (ICML). 2026.

AuthorsOpenReview ↗

Response to reviewer VQ2f (2/3)

Weakness 2: Limited expansion of demonstrated task capabilities. The method offers a useful simplification and greater flexibility, but the experiments do not establish acceleration capabilities beyond the task regimes already explored by prior work. Evidence on tighter-tolerance manipulation would strengthen the empirical contribution.

Question 2: How well does cosine inconsistency capture criticality when precise execution requires directionally consistent motion, such as tight-tolerance insertion? Evaluation or failure analysis on such tasks would help establish the limits of the proposed signal.

Response: Thank you for the suggestion on tight-tolerance manipulation. It prompted us to add an evaluation on an insertion task and to distinguish more carefully between the overall direction of motion in a task and the local direction changes within a predicted action chunk.

1) The new experiment covers insertion along a fixed direction after alignment.

In Appendix E.6 of the revised manuscript, we add a tight-tolerance insertion task built on the Robomimic Square task. We thicken the square peg so that the clearance per side between the nut and the peg shrinks from 6.75 mm to 1.75 mm, and we keep all other settings unchanged. We train a Flow Matching policy on 50 demonstrations of this task. To isolate the tight-tolerance phase, every evaluation rollout starts with the nut already grasped, and all commanded rates start from the same 50 grasped states. For commanded rates up to 3×, FreeSpeed modulates the execution speed while maintaining success, with success rates between 90% and 96% compared with 92% for the reference. At 4×, the success rate drops markedly. Table A16 gives the full results, and the table below summarizes the results of FreeSpeed at each commanded rate, with 50 rollouts per condition.

Command1× (reference)1.5×2×3×0.5×0.3×
Success rate92%96%96%94%92%90%
Realized rate1.00×1.34×1.64×1.94×0.65×0.34×

The analysis of the demonstrations in Figure A6 shows that the alignment phase involves frequent direction corrections, which lead to higher cosine inconsistency and more conservative rescaling. After alignment, the insertion proceeds mainly downward along a fixed direction, so cosine inconsistency decreases and the rescaling factor rises again. In this task, alignment and insertion thus exhibit different local action characteristics, and tight-tolerance manipulation still leaves room for speed modulation.

The insertion can be accelerated because the outcome of this task is largely decided when the nut is aligned with the peg and placed over it. After that, the downward motion is no longer critical, and pushing the nut down completes the task. This task therefore does not fully represent the case you describe, in which the motion is both straight and critical.

To examine that case, we also built a pure insertion task in which the insertion itself is the critical phase (point 2 below). On such a task, however, the base policy succeeds too rarely for a comparison of success rates to be statistically meaningful, which is consistent with the low success rates commonly reported for insertion tasks in prior work. We therefore analyze it through its demonstrations.

AuthorsOpenReview ↗

Response to reviewer VQ2f (3/3)

Response to Weakness 2 and Question 2 (continued):

2) A consistent overall direction does not imply consistent directions within an action chunk.

During an insertion that requires continual contact corrections, the overall motion can advance along the insertion axis while still containing local lateral adjustments or retreats. When such corrections appear in the translation increments of a predicted action chunk, cosine inconsistency can be high, which drives the rescaling factor down. Whether the signal is effective therefore depends on the local actions that the policy actually predicts, not merely on whether the task geometry follows a straight line.

To examine such a task, we built a peg insertion scene based on Factory [1] in which the peg advances only under a sustained pushing force of 15 N. We collected 50 teleoperated demonstrations in this scene and computed their cosine inconsistency.

Figure A7 of the revised manuscript presents the results. Its main message is that an insertion that is straight in geometry is not straight in the action space. Although the peg moves along a single axis, the commanded increments keep changing direction, because the insertion is regulated by contact force. The commanded pose must lead the actual pose to generate the insertion force, and it retreats once the peg is seated. Cosine inconsistency in panel (b) is therefore high throughout the task, and it rises further as the contact force in panel (d) builds up from about 6 N at the hole rim to about 15 N during the insertion. The gate thus treats the insertion as a critical phase, and the median rescaling factor in panel (c) stays at 1 in all phases. In other words, when precise execution depends on contact force, its criticality appears as direction changes in the actions, which is exactly what cosine inconsistency measures. The force-aware policy ForceVLA [2] succeeds in only 1 of 20 rollouts on this task, so we did not evaluate rate modulation on it.

3) Critical interactions that are directionally consistent yet speed-sensitive are a limitation of the signal.

The case you describe can indeed occur. If the predicted actions keep a consistent direction during a critical interaction, cosine inconsistency remains low even though this phase requires careful execution. For example, when the policy completes a grasp decisively along a single direction without any correction, enlarging the step length may still cause an overshoot or a failed grasp, and the gate may not detect this risk in advance.

Cosine inconsistency therefore modulates speed on the basis of local changes in action direction, and it cannot guarantee that every speed-sensitive critical interaction is identified. We now state this boundary explicitly in the Limitations and use the new insertion experiments to illustrate both the action characteristics that the signal captures and its limits.

Revision: Appendix E.6 now reports the tight-tolerance insertion experiment in simulation. Figure A6 shows how cosine inconsistency and the rescaling factor vary across task phases on the demonstrations, and Table A16 reports the success rates and realized rates at each commanded rate. Figure A7 shows how cosine inconsistency, the rescaling factor, and the contact force vary across task phases on the demonstrations of the force-controlled press-fit task. The Limitations now state the scope of the signal.

References:

[1] Y. Narang, et al. "Factory: Fast contact for robotic assembly." Robotics: Science and Systems (RSS). 2022.

[2] J. Yu, et al. "ForceVLA: Enhancing VLA models with a force-aware MoE for contact-rich manipulation." Advances in Neural Information Processing Systems (NeurIPS). 2025.

Reviewer_oJazOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper proposes a training-free post-processor that changes the execution speed of a frozen action-chunking policy with delta end-effector actions. At each replan, the predicted chunk is resampled with stride ρ\rho (skipping actions to accelerate, splitting them to decelerate), and the translation and rotation increments are rescaled by a factor that moves from a speed-matching value toward 1 as the chunk's cosine inconsistency (the largest direction change between consecutive increments) grows. One endpoint table and λ=50\lambda = 50 are shared across all policies and environments. The motivation is that speed changes shift the observation distribution mainly in high-curvature, contact-critical phases. Validation covers flow-matching policies on 10 LIBERO-90 tasks and four real-robot tasks, released π0.5\pi_{0.5} and Fast-WAM checkpoints on the four LIBERO suites, Diffusion Policy on Robomimic, and π0.5\pi_{0.5} on CALVIN, with commands from 0.2× to 4×, fixed or changing per replan. FreeSpeed largely preserves reference success where vanilla resampling and naive rescaling degrade, while realized rates stay well below the commands (e.g., 1.75× for a 4× command on the robot).

Critical strengths
  1. FreeSpeed generally achieves better success rates than the baselines at similar actual speeds. For example, 89% vs. 80% on LIBERO-90, 91.2% vs. 77.5% on the real robot, and 64.9% vs. 59.8% on CALVIN. The π0.5 results show a similar trend, though RoboMimic is an exception. I find these results fairly convincing.
  2. The method requires no retraining or policy changes and works across several policy families and benchmarks. It also supports slowing down and changing speeds during execution, which makes it useful in practice.
Critical weaknesses
  1. It's unclear how much the inconsistency gate actually helps. In App. E.2, a fixed factor of 1.2 achieves the same 95.3% success rate as FreeSpeed, with only slightly lower speed (1.30× vs. 1.38×). More comparisons at higher speeds and on the real robot would help clarify its contribution.
  2. The actual speed differs quite a bit from the requested speed (e.g., 4× only reaches 1.75×, and 0.2× reaches 0.40×). Also, speedups are measured by control-step ratios rather than wall-clock time, so the actual runtime benefit remains unclear.
Questions for authors
  1. Can you extend the fixed-factor baseline to 3×/4×, deceleration, π0.5\pi_{0.5}, and the real robot? How does the inconsistency gate compare with simpler signals like action magnitude?
  2. Is robot inference synchronous or asynchronous? What are the actual wall-clock speedups, and can π0.5\pi_{0.5} maintain real-time control with only 2 executed steps per replan?
  3. How were fc,λ=50f_c, \lambda=50, and fs=0.9ρf_s=0.9\rho chosen? Which task is shown in Fig. 7b,c? Were any parameters tuned on the evaluation tasks?
  4. Could the improvement simply come from more frequent replanning? How would vanilla perform with the same number of executed steps per replan?
  5. Can you provide success vs. actual speed curves with confidence intervals, especially for RoboMimic? Can the method better match the requested speed?
11 comments
AuthorsOpenReview ↗

Response to reviewer oJaz (1/7)

We sincerely thank you for your careful and constructive comments. We are glad that you are convinced by the success rates of FreeSpeed relative to the baselines at comparable realized rates, and that you recognize the practicality of a method that requires no retraining, applies to multiple policy families and benchmarks, and supports both deceleration and rate changes during execution. Below we respond to each weakness and question, and we hope that these responses address your concerns. A revised manuscript has also been uploaded, with all additions and changes marked in blue.

Weakness 1: It's unclear how much the inconsistency gate actually helps. In App. E.2, a fixed factor of 1.2 achieves the same 95.3% success rate as FreeSpeed, with only slightly lower speed (1.30× vs. 1.38×). More comparisons at higher speeds and on the real robot would help clarify its contribution.

Question 1: Can you extend the fixed-factor baseline to 3×/4×, deceleration, π0.5, and the real robot? How does the inconsistency gate compare with simpler signals like action magnitude?

Response: Thank you for your careful reading. Weakness 1 and Question 1 concern the same point, so we address them together in three parts.

1) Extension of the fixed-factor baseline. Compared with the fixed-factor baseline, the main advantage of the gate is that a single setting stays close to the best fixed-factor trade-off on every task, which no single fixed factor achieves. This matters because the best constant factor is known only a posteriori. It is task-dependent and cannot be known in advance, so it has to be found by sweeping the factor on each task, whereas FreeSpeed uses one setting for all tasks. The existing results in Appendix E.2 already support this. In Figure A3, the constant factor with the highest joint score of success rate and realized rate is 1.3 on Task 34, 1.8 on Task 48, and 1.4 on Task 86. Moreover, the factor of 1.8 that is best for Task 48 reduces the success rate on Task 86 from the 92% of the 1× reference to 54%. Similarly, a factor such as 1.2, as mentioned in Weakness 1, is not optimal for all tasks.

To further verify the above, we evaluated the constant-factor sweep on two additional LIBERO-90 tasks, Tasks 18 and 60. The table below adds them to Table A11 (2× command, 50 rollouts per task and condition, constant factors from 1.0 to 2.0). The last row, in bold, reports the joint score used in Figure A3, the product of the success rate S∈[0,1]S \in [0, 1] and the realized rate r^\hat{r}, averaged over the five tasks. The constant factor with the highest joint score becomes 1.4.

TaskReference1.01.11.21.31.41.51.61.71.81.92.0FreeSpeed
18SR (%)98969296969494929090868298
Rate (×)1.001.061.131.231.251.351.441.491.571.601.561.531.37
34SR (%)9810010010010096928470623040100
Rate (×)1.001.191.261.321.351.381.401.130.981.060.991.181.36
48SR (%)9210010094929296888888724698
Rate (×)1.001.051.171.271.311.331.371.451.481.541.451.501.35
60SR (%)1006888969694867670745260100
Rate (×)1.001.091.181.251.301.351.341.321.201.271.251.381.34
86SR (%)92929492908878565654404088
Rate (×)1.001.171.221.311.301.421.491.531.451.671.471.451.43
MeanSR (%)96.091.294.895.694.892.889.279.274.873.656.053.696.8
Rate (×)1.001.111.191.271.301.361.411.381.331.431.341.411.37
Joint score0.961.011.131.221.231.271.251.091.011.060.780.761.33

To examine this more broadly, you suggest extending the fixed-factor baseline to the 3× and 4× commands, deceleration, π0.5\pi_{0.5}, and the real robot. These experiments are extensive and are currently running, and we will post the results here as soon as they are available. All the other weaknesses and questions you raised are answered below, and the corresponding revisions have been made.

AuthorsOpenReview ↗

Response to reviewer oJaz (2/7)

Response to Weakness 1 and Question 1 (continued):

2) Cosine inconsistency versus a simpler signal. Following your suggestion, we compared cosine inconsistency against action magnitude, measured as the translation magnitude of each predicted action averaged over the chunk. With this signal, a smaller average magnitude leads to more conservative rescaling.

For a fair comparison, the action-magnitude signal should accelerate at least as much as FreeSpeed, so we calibrate it on the demonstrations. It uses the same factor endpoints and the same mapping as FreeSpeed, and its reference value is the median of this chunk-averaged magnitude over the demonstrations. With this setting, its mean rescaling factor over the demonstration chunks under the 2× command is comparable to that of FreeSpeed. The table below reports the success rates (%) of Flow Matching on the ten LIBERO-90 tasks under the 2× and 0.5× commands, with 50 rollouts for each task and condition. Bold marks the mean success rate of the reference and the higher mean success rate under each command. The results of the reference and FreeSpeed are taken from Table 2, and the per-task realized rates are given in Table A13.

Under the 2× command, FreeSpeed reaches a mean success rate of 88.6% at a realized rate of 1.35×, on par with the 88.8% of the 1× reference. In contrast, the action-magnitude signal reaches a mean success rate of 63.8% at a realized rate of 1.21×, and FreeSpeed has the higher success rate on all ten tasks. Under the 0.5× command, FreeSpeed reaches a mean success rate of 89.8% and the action-magnitude signal reaches 86.6%, both at a realized rate of 0.63×, a smaller gap in success than under the 2× command. This is consistent with the general pattern that deceleration affects success less than acceleration does.

CommandMethod2324086184860346569Mean
1×Reference96789092989210098648088.8
2×FreeSpeed887688889898100100569488.6
Action magnitude5072586486947858146463.8
0.5×FreeSpeed9270949098989496788889.8
Action magnitude7270889294969692848286.6

3) Why we chose cosine inconsistency. We chose cosine inconsistency as the gating signal because it captures, directly from the action chunk that the policy predicts, the local direction changes that often accompany critical behaviors such as alignment and fine adjustment. It therefore indicates when rescaling should be conservative, and it requires no additional phase labels or training.

We hope that this addresses your concern, and we would be glad to clarify further if needed.

Revision: Section 5.7 now reports the comparison with the action-magnitude signal, and Appendix E.2 now includes a new paragraph and Table A13 with the per-task results.

AuthorsOpenReview ↗

Response to reviewer oJaz (3/7)

Weakness 2: The actual speed differs quite a bit from the requested speed (e.g., 4× only reaches 1.75×, and 0.2× reaches 0.40×). Also, speedups are measured by control-step ratios rather than wall-clock time, so the actual runtime benefit remains unclear.

Question 2: Is robot inference synchronous or asynchronous? What are the actual wall-clock speedups, and can π0.5 maintain real-time control with only 2 executed steps per replan?

Response: Thank you for raising these points. Your observations are accurate. The realized rate is indeed lower than the commanded rate, which is determined by the mechanism of FreeSpeed. Weakness 2 and Question 2 concern the gap between the commanded rate and the realized rate, as well as the actual runtime benefit, so we address them together in three parts. Below, we first explain the actual runtime and the inference mode, then the gap between the commanded rate and the realized rate, and finally the real-time control of π0.5\pi_{0.5} with 2 executed steps per replan.

1) Actual runtime and inference mode.

Real robot. Inference is synchronous. The control loop runs at a fixed 30 Hz, and while the next chunk is being computed, the controller sends hold commands. These hold steps are counted as steps, so the reported realized rate is the wall-clock rate, with inference waiting included. The table below gives the mean completion time of FreeSpeed on the four real-robot tasks in seconds, obtained by dividing the step counts in Table A14 by 30 Hz.

Task1× reference2×3×4×0.5×0.3×0.2×
Peach10.87.86.86.714.422.226.9
Tennis17.212.211.09.724.931.540.0
Stack Cube16.19.89.08.124.936.542.6
Pour Almond29.221.719.117.644.669.976.9
Mean of four tasks18.312.911.510.527.240.046.6

Simulation. Inference is also synchronous, but the simulator advances only when an action is executed, so the inference time is not counted. Each step corresponds to 50 ms of simulated time, so the step count gives the execution time of the task, which is what FreeSpeed changes. We therefore use the ratio of step counts over successful rollouts as the realized rate. This measure depends only on the method and not on the inference hardware or the deployment, which keeps the comparison across methods fair and reproducible. SuP (Wu et al., 2026) and TempoVLA (Jing et al., 2026) use the same measure (their Table 2 and Table 3), so the comparison with them in Figure 7(a) is direct.

2) Realized rate versus commanded rate. The realized rate falls below the commanded rate for two reasons. First, the commanded rate is the rate of FreeSpeed at its most aggressive, not its average rate, because fine-grained phases are deliberately kept close to the original speed. Second, closed-loop execution adds corrective motions, which take additional steps. The same gap appears in other methods, including those that retrain the policy. Under the 2× command, the speed-conditioned policy of TempoVLA (Jing et al., 2026) realizes 1.56× (their Table 3), and test-time 2× downsampling realizes 1.72× in SuP (Wu et al., 2026) (their Table 2) and 1.6× to 1.7× in DemoSpeedup (Guo et al., 2025) (their Table 1). These methods also treat task phases differently. DemoSpeedup downsamples the demonstrations non-uniformly under entropy guidance before retraining, and SuP learns a scheduler that selects the downsampling rate during execution. As we also discuss in the Limitations (Section 6, Appendix F), the achievable speedup is ultimately bounded by the task itself, since it depends on how much of the task can be executed through directionally consistent motion.

(Continued in the next comment.)

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of the rebuttal and the major changes in the revised manuscript

Dear Reviewers,

We sincerely thank you for the time and care you devoted to reviewing our manuscript. Your constructive comments have helped us improve both the rigor and the clarity of this work.

We appreciate that the reviewers recognize the strengths of FreeSpeed: "The paper's perspective is interesting, the method is simple, and the experiment results demonstrate the effectiveness of the proposed method." (Reviewer B9W9), "A practical contribution with a coherent motivation." and "provides a more flexible control interface." (Reviewer VQ2f), "which makes it useful in practice." and "I find these results fairly convincing." (Reviewer oJaz), and "The paper is well-written, with ideas and experiment results demonstrated clearly." (Reviewer B9W9).

The reviewers also raised valuable concerns and questions. We have addressed each of them in the individual responses, with additional experiments and analyses in support. The revised manuscript has been uploaded, with all additions and changes marked in blue. Owing to the page limit, the key results are reported in the main text and the complete results in the appendix. The major changes are summarized below.

  1. Comparison experiment. As suggested, we add a comparison with action magnitude as the gating signal, evaluated on the ten LIBERO-90 tasks under the 2× and 0.5× commands (Section 5.7, Appendix E.2, Table A13).

  2. Wall-clock runtime. We report the wall-clock completion time on the real robot (Table 3, Table A14) and estimate the wall-clock rate of π0.5\pi_{0.5} with inference waiting included (Appendix E.1, Table A10).

  3. Replanning frequency. We add a controlled experiment that varies the number of actions executed per replan (Appendix E.2, Table A12).

  4. Confidence intervals. We report success rates and realized rates with 95% confidence intervals for all four policy families (Figure 8).

  5. Rescaling across task phases. We analyze cosine inconsistency and the rescaling factor in each task phase, together with its realized rate (Section 5.2, Appendix E.5).

  6. Tight-tolerance insertion. We add a tight-tolerance insertion task and examine the scope of the signal on a press-fit peg insertion (Section 5.3, Appendix E.6).

  7. Paired analysis of task success. We compare the reference policy and FreeSpeed on identical initial states (Appendix E.7, Figure A8).

  8. Method and experimental details. We define the resampling for acceleration and deceleration with worked examples (Section 4.2, Table A5), and specify the task and results of the sensitivity analysis (Section 5.7, Figure 7).

  9. Limitations. We discuss the scope of the signal and the additional policy inferences required under acceleration (Section 6, Appendix F).

Thank you again for your valuable feedback. We look forward to further discussion.

Best regards,

Authors of Submission 13981

AuthorsOpenReview ↗

Summary of the Rebuttal Discussion

Dear Reviewers and Area Chair,

Thank you for your time and consideration.

Reviewers’ feedback. We appreciate the reviewers’ consistently positive assessments, with all three initially rating our paper 3. During the rebuttal, we provided additional experiments and clarifications in response to the reviewers’ concerns, and Reviewer 1 raised the soundness score. Reviewers 2 and 3 also expressed their continued support for acceptance.

Paper contributions. Our key finding is that directional inconsistency within predicted action chunks provides a useful signal of task-phase criticality. FreeSpeed uses this signal to enable both acceleration and deceleration of frozen policies without additional training, while preserving task success. This offers a simple way to give pretrained robot policies online speed control, allowing execution speed to adapt to changing deployment needs.

We are grateful for the reviewers’ constructive feedback, which has helped us present our work more clearly. Thank you again for your careful evaluation of our work.

Authors of Submission 13981

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.