FreeSpeed · Training-Free Speed Control

(a) Demonstration retiming compared with FreeSpeed's test-time resampling and rescaling of predicted action chunks. (b) Success rates of FreeSpeed and vanilla resampling on three policies under speedup and slowdown commands. — **FreeSpeed modulates the execution speed of a frozen policy.** (a) Methods based on demonstration retiming resample demonstrations and train or fine-tune the policy, while FreeSpeed resamples and rescales predicted action chunks at test time. (b) Success rates of FreeSpeed and vanilla resampling on three policies under speedup and slowdown commands.

Overview

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 0.22× to 2.53× among settings that preserve per-task success. Across four real-world manipulation tasks, FreeSpeed achieves an average success rate of 94.0%, matching the frozen policy's 93.8%, while realizing execution rates from 0.38× to 1.97×.

3 policiesFrozen π_0.5, Fast-WAM, and task-specific Flow Matching

50 + 4 tasksLIBERO simulation tasks and real-robot tasks

0.22×–2.53×Realized rates in simulation among settings that preserve per-task success

94.0%Real-robot mean success over six commands, vs. 93.8% for the frozen policy

Related work

Speed-control capabilities of the compared methods

Method	No rate training	Unmodified demos	Unchanged controller	User-set rate	Slow-motion	In-episode rate
SuP	○	○	○	○	○	○
SpeedTuning	○	●	●	○	○	○
TempoVLA	○	○	●	●	●	●
SAIL	○	○	○	○	○	○
RACE	●	○	○	○	○	○
DemoSpeedup	○	○	○	○	○	○
AutoSpeed	○	●	○	○	●	○
SpeedAug	○	○	●	○	○	○
FreeSpeed (ours)	●	●	●	●	●	●

A filled dot indicates that the capability is demonstrated in the corresponding paper; an open dot indicates that it is not demonstrated.

Method

FreeSpeed at one replan: the frozen policy predicts an action chunk; translation increments give a cosine inconsistency score; the rate command determines temporal resampling; together they set the rescaling factor. — **FreeSpeed at one replan.** The frozen policy predicts an action chunk. Its translation increments provide a cosine inconsistency score, and the rate command determines temporal resampling. Together, the score and command determine the factor used to rescale the resampled translation and rotation increments.

1Resample

At each replan, FreeSpeed resamples the predicted chunk of incremental (delta) actions on a temporal grid with stride ρ_t. Acceleration skips predicted actions, while deceleration splits each action into sub-actions that preserve its direction.

2Measure inconsistency

Cosine inconsistency measures the change in direction between consecutive translation increments. The inconsistency of a chunk, I_cos, is the maximum over its H−1 adjacent increment pairs.

3Rescale

f_t = f_c + (f_s − f_c) exp(−λI_cos)

The factor scales translation and rotation increments; gripper commands are unchanged. It equals f_s when I_cos is zero and approaches f_c as I_cos increases.

Signal

Cosine inconsistency tracks the task phase

Segments with high I_cos correspond to decisive task phases, identified in these examples as grasping and placement in pick-and-place tasks. Under acceleration, the factor is large when the chunk follows a straight trajectory and decreases toward 1.0 as CI increases. Under deceleration, the factor is close to the commanded stride when CI is small and increases as CI grows. In both cases, increasing CI brings the executed step lengths closer to those originally predicted by the policy.

**The signal and the response.** (a) Cosine inconsistency along human demonstrations of five LIBERO-90 tasks, with peaks during grasping and placement. (b) The mapping from cosine inconsistency to the rescaling factor under the 2× and 0.5× commands.

Experiments

Evaluation environments: 50 simulation tasks from LIBERO and four real-robot tasks. — **Evaluation environments.** 50 simulation tasks from LIBERO and four real-robot tasks.

Protocol

Policies and tasks

We evaluate three policy architectures: VLA, WAM, and task-specific flow matching. Checkpoints of π_0.5 and Fast-WAM were pretrained on 40 tasks spanning LIBERO-Spatial, Object, Goal, and Long and are kept frozen. Task-specific flow-matching policies are trained on 10 LIBERO-90 tasks and 4 real-robot tasks. On the real robot, a Rokae Helios series humanoid executes the action at 30 Hz.

Baselines

Compared execution rules

Vanilla resampling sums the skipped action increments into the retained action. w/o CR combines resampling with naive rescaling by the commanded stride, f_t = ρ_t, as an ablation without CI-based rescaling. The unchanged 1× policy is the reference. Under deceleration, vanilla resampling and w/o CR are equivalent, so only w/o CR is reported.

Simulation

π_0.5 and Fast-WAM on four LIBERO suites

At a 2× command, FreeSpeed maintains the reference success rates for both policies, whereas vanilla resampling reduces success by 9.2 and 12.7 percentage points and w/o CR by 8.0 and 12.3 percentage points, respectively. Under deceleration, FreeSpeed outperforms w/o CR at every slowdown stride on both policies. Across individual tasks from both policies, the realized rates that preserve each task's own 1× success rate range from 0.22× to 2.53×.

Success rates and realized execution rates of π0.5 and Fast-WAM on the four LIBERO suites under speedup and slowdown commands. — **Simulation comparison for π_0.5 and Fast-WAM.** Bars show success rates on the left axes, and triangular markers show realized execution rates on the right axes. The rate axes are inverted in the slowdown panels.

Simulation

Flow Matching on 10 LIBERO-90 tasks

Across the four non-reference strides, FreeSpeed maintains or exceeds the 1× reference's mean success rate of 89%, with realized rates from 0.348× to 1.431× across individual tasks.

	Success (%)			Rate (×)
Cmd	w/o CR	Vanilla	Ours	w/o CR	Vanilla	Ours
1× ref.	89			1.000
1.5×	73	80	90	1.242	1.280	1.221
2×	46	61	89	1.435	1.483	1.353
0.5×	84	—	90	0.570	—	0.626
0.3×	74	—	90	0.328	—	0.402

Mean over 10 tasks, 50 rollouts per task. Rate averages per-task realized rates, each computed over trials in which both the evaluated method and the reference succeed.

Real robot

Four real-robot tasks

Averaged over the six non-reference strides, FreeSpeed achieves 94.0% mean success, compared with the reference's 93.8%, with realized execution rates from 0.38× to 1.97× across individual tasks.

	Success (%)			Rate (×)
Cmd	w/o CR	Vanilla	Ours	w/o CR	Vanilla	Ours
1× ref.	93.8			1.00
2×	72.5	77.5	97.5	1.66	1.69	1.44
3×	37.5	38.8	95.0	1.86	1.93	1.62
4×	25.0	11.2	91.2	1.93	2.18	1.75
0.5×	82.5	—	96.2	0.57	—	0.68
0.3×	66.2	—	93.8	0.37	—	0.47
0.2×	55.0	—	90.0	0.25	—	0.40

Mean over Peach on Plate, Tennis in Can, Stack Cube, and Pour Almond.

Varying rate

Rate that varies during execution

FreeSpeed maintains task success under stride randomization across policy inferences and factor randomization within its bounds. (A) A new stride is sampled from the full command set at each replan. (B) The stride remains fixed, while the rescaling factor is sampled between f_c and f_t; success rates remain within 2 percentage points of the fixed-stride rule.

A: the stride varies
Policy	Ref. succ.	Succ.	Rate
Flow Matching	88	90	0.63
π_0.5	97	97	0.54

B: the factor varies · success (%) / rate (×)
Policy	Factor	0.3×	0.5×	1.5×	2×
FM	fixed	90 / 0.40	90 / 0.63	90 / 1.22	89 / 1.35
FM	varying	89 / 0.46	91 / 0.68	92 / 1.20	91 / 1.28
π_0.5	fixed	97 / 0.39	98 / 0.57	97 / 1.21	97 / 1.34
π_0.5	varying	97 / 0.42	98 / 0.61	96 / 1.13	97 / 1.21

FM is evaluated on the 10 LIBERO-90 tasks, and π_0.5 on all 40 LIBERO tasks.

Real-robot rollouts

Select a task and a commanded rate to compare the 1× reference, w/o CR, vanilla resampling, and FreeSpeed side by side. Select Variable to see the rate change within an episode.

Command4×

Analysis

Observation distribution

FreeSpeed keeps the policy in distribution

On LIBERO-90 Task 60 with Flow Matching at 2×, FreeSpeed remains close to the 1× reference and training data, with a median distance of 1.09 times the reference median, compared with 1.79 for vanilla resampling and 1.68 for w/o CR. Under the two baselines, higher inconsistency is associated with larger subsequent distances, whereas FreeSpeed stays close to the reference across the inconsistency range.

Observation distribution shift under the 2× command: FreeSpeed remains close to the 1× reference, while the baselines deviate further from the training data, particularly after chunks with high cosine inconsistency. — **Observation distribution shift under the 2× command.** FreeSpeed remains close to the 1× reference in (a) and (b), while the baselines show larger deviations from the training data, particularly after executing chunks with high cosine inconsistency in (c).

Comparison and sensitivity

Prior methods and the two constants

(a) Success change against realized rate for FreeSpeed, vanilla resampling, SuP, and TempoVLA. (b, c) The decay rate lambda and the factor f s swept in turn. — **Comparison with published methods and sensitivity to the two constants.** (a) Success change against realized rate. (b, c) The decay rate λ and the factor f_s swept in turn.

On LIBERO, FreeSpeed matches SuP and TempoVLA in success-rate change and realized rate while providing a training-free, plug-and-play solution. Vanilla resampling can achieve higher realized rates, but at the cost of a substantial reduction in success rate. Changes are reported rather than absolute success rates because evaluation settings differ.

The decay rate λ and the factor f_s are varied in turn at the 2× command, with 50 rollouts per setting, showing a broad range of effective settings for both. Shading indicates settings whose success rates remain within ten percentage points of the best observed result.

Conclusion

FreeSpeed enables test-time execution-speed control for frozen action-chunking policies. It computes cosine inconsistency between adjacent actions within each predicted chunk and uses this signal to adapt rescaling under fixed or time-varying commands. The main limitation of FreeSpeed is that its acceleration ceiling is set by how much free motion a task contains.

BibTeX

@misc{hu2026freespeed,
  title  = {FreeSpeed: Training-Free Speed Control
            for Generative Robot Policies},
  author = {Hu, Yuxuan and Shan, Shilin and Wang, Qiheng and Yang, Jinghan
            and Fan, Junqiao and Wan, Hao and Yang, Jianfei},
  year   = {2026}
}

FreeSpeedTraining-Free Speed Control for Generative Robot Policies