FreeSpeed: Training-Free Speed Control for Generative Robot Policies

Yuxuan Hu; Shilin Shan; Qiheng Wang; Jinghan Yang; Junqiao Fan; Hao Wan; Jianfei Yang

cs.ROOct 5, 2026

FreeSpeed: Training-Free Speed Control for Generative Robot Policies

Authors: Yuxuan Hu, Shilin Shan, Qiheng Wang, Jinghan Yang, Junqiao Fan, Hao Wan, Jianfei Yang

Organizations: MARS Lab, Nanyang Technological University · ROKAE Robotics

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 observe that the directional inconsistency of action chunks reflects task-phase criticality, indicating how aggressively action step lengths can be modified while preserving task success. Based on this observation, 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 94.0%, matching the frozen policy's 93.8%, while realizing execution rates from 0.38x to 1.97x.

Figures & tables

Figure 1: 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.

	No rate	Unmodified	Unchanged	User-set	Slow-	In-episode
	training	demos	controller	rate	motion	rate
SuP ( Wu et al., 2026 )	$\circ$	$\circ$	$\circ$	$\circ$	$\circ$	$\circ$
SpeedTuning ( Yuan et al., 2025 )	$\circ$	$\bullet$	$\bullet$	$\circ$	$\circ$	$\circ$
TempoVLA ( Jing et al., 2026 )	$\circ$	$\circ$	$\bullet$	$\bullet$	$\bullet$	$\bullet$
SAIL ( Arachchige et al., 2025 )	$\circ$	$\circ$	$\circ$	$\circ$	$\circ$	$\circ$
RACE ( Kim et al., 2026 )	$\bullet$	$\circ$	$\circ$	$\circ$	$\circ$	$\circ$

Table 1: Speed-control capabilities of the compared methods. A filled dot indicates that the capability is demonstrated in the corresponding paper; an open dot indicates that it is not demonstrated.

Figure 2: 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.

Figure 3: Evaluation environments. 50 simulation tasks from LIBERO and four real-robot tasks.

Figure 4: 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\times

and

0.5\times

commands.

Type	Task	$1\times$	$1.5\times$			$2\times$			$0.5\times$		$0.3\times$
		Ref.	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	Free [-2pt]Speed	w/o [-2pt]CR	Free [-2pt]Speed
Structure constrained	2	$\mathbf{96}$	$72$	$82$	$\mathbf{94}$	$28$	$36$	$\mathbf{88}$	$74$	$\mathbf{92}$	$46$	$\mathbf{90}$
	32	$\mathbf{78}$	$\mathbf{76}$	$66$	$\mathbf{76}$	$52$	$50$	$\mathbf{76}$	$58$	$\mathbf{70}$	$66$	$\mathbf{76}$
	40	$\mathbf{90}$	$70$	$\mathbf{90}$	$88$	$78$	$78$	$\mathbf{88}$	$\mathbf{100}$	$94$	$90$	$\mathbf{96}$
	86	$\mathbf{92}$	$70$	$84$	$\mathbf{92}$	$40$	$64$	$\mathbf{88}$	$82$	$\mathbf{90}$	$\mathbf{86}$	$\mathbf{86}$
Bounded receptacle	18	$\mathbf{98}$	$92$	$96$	$\mathbf{98}$	$82$	$96$	$\mathbf{98}$	$92$	$\mathbf{98}$	$\mathbf{96}$	$\mathbf{96}$

Table 2: Fixed-stride execution on 10 LIBERO-90 tasks with Flow Matching. Success rates are reported in %, with task indices defined in Table A2 . ‘Rate’ is relative to the

1\times

reference and averages per-task realized rates, each computed over trials in which both the evaluated method and the reference succeed. Bold indicates the highest success rate under each command. Completion steps are reported in Table A5 .

	$1\times$	$2\times$			$3\times$			$4\times$			$0.5\times$		$0.3\times$		$0.2\times$
	Ref.	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	Free [-2pt]Speed	w/o [-2pt]CR	Free [-2pt]Speed	w/o [-2pt]CR	Free [-2pt]Speed
Peach on Plate	$100$	$70$	$90$	$\mathbf{100}$	$50$	$55$	$\mathbf{100}$	$45$	$0$	$\mathbf{95}$	$70$	$\mathbf{100}$	$65$	$\mathbf{100}$	$20$	$\mathbf{95}$
Tennis in Can	$90$	$90$	$60$	$\mathbf{95}$	$45$	$30$	$\mathbf{90}$	$0$	$0$	$\mathbf{85}$	$80$	$\mathbf{95}$	$70$	$\mathbf{90}$	$65$	$\mathbf{95}$
Stack cube	$90$	$60$	$75$	$\mathbf{95}$	$0$	$20$	$\mathbf{90}$	$5$	$0$	$\mathbf{90}$	$\mathbf{95}$	$90$	$65$	$\mathbf{95}$	$\mathbf{75}$	$\mathbf{75}$
Pour almond	$95$	$70$	$85$	$\mathbf{100}$	$55$	$50$	$\mathbf{100}$	$50$	$45$	$\mathbf{95}$	$85$	$\mathbf{100}$	$65$	$\mathbf{90}$	$60$	$\mathbf{95}$
Mean	$93.8$	$72.5$	$77.5$	$\mathbf{97.5}$	$37.5$	$38.8$	$\mathbf{95.0}$	$25.0$	$11.2$	$\mathbf{91.2}$	$82.5$	$\mathbf{96.2}$	$66.2$	$\mathbf{93.8}$	$55.0$	$\mathbf{90.0}$

Table 3: Fixed-stride execution on four real-robot tasks. Per-task success rates are reported in %. Bold indicates the highest success rate under each command. Per-task completion steps and realized rates are reported in Table A10 .

Figure 5: Simulation comparison for $\pi_{0.5}$ 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.

Figure 6: Observation distribution shift under the $2\times$ command. FreeSpeed remains close to the $1\times$ 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).

		A: the stride varies		B: the factor varies
		FM	$\pi_{0.5}$	FM				$\pi_{0.5}$
		dynamic	dynamic	$0.3\times$	$0.5\times$	$1.5\times$	$2\times$	$0.3\times$	$0.5\times$	$1.5\times$	$2\times$
FreeSpeed	Succ.	–	–	$90$	$90$	$90$	$89$	$97$	$98$	$97$	$97$
	Rate	–	–	$0.40$	$0.63$	$1.22$	$1.35$	$0.39$	$0.57$	$1.21$	$1.34$
varying	Succ.	$90$	$97$	$89$	$91$	$92$	$91$	$97$	$98$	$96$	$97$
	Rate	$0.63$	$0.54$	$0.46$	$0.68$	$1.20$	$1.28$	$0.42$	$0.61$	$1.13$	$1.21$

Table 4: Within-episode stride and factor randomization. FM is evaluated on the

10

tasks in Table 2 , and

\pi_{0.5}

on all

40

tasks in Figure 5 . (A) A new stride is sampled at each replan. (B) The stride remains fixed, while the rescaling factor is sampled between

f_{\mathrm{c}}

and

f_{t}

Figure 7: Comparison with published methods and sensitivity to the two constants. (a) Success change against realized rate. (b, c) The decay rate

\lambda

and the factor

f_{\mathrm{s}}

swept in turn.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1: The real-robot workstation. The experiments use the robot’s left arm, shown on the right in the image. This frame is taken from a $1\times$ rollout of the tennis-ball task, with the ball in the gripper and two cans on the table.

Item	Value
Robot	Rokae Helios series humanoid
Arm	7 DoF
End effector	Daimon gripper
robot command rate	$30$ Hz
Action interface	Cartesian end-effector increments
Head camera	Orbbec, $640\times 480$ , $30$ fps

Appendix

Table A1: Real-robot hardware and demonstration collection.

Type	Task	Language instruction
Structure constrained	2	put the black bowl in the top drawer of the cabinet
	32	put the ketchup in the top drawer of the cabinet
	40	put the frying pan on the cabinet shelf
	86	pick up the book in the middle and place it on the cabinet shelf
Bounded receptacle	18	put the frying pan on the stove
	48	pick up the ketchup and put it in the basket

Appendix

Table A2: The 10 LIBERO-90 tasks used for the flow-matching evaluation. Each task is listed with its LIBERO-90 index and corresponding language instruction.

	Flow Matching	$\pi_{0.5}$	Fast-WAM
Source	trained in this work	released checkpoint	released checkpoint
Observation history	$8$ frames	$1$ frame	$1$ frame
Steps executed per replan	$12$	$8$	$10$
Evaluation tasks	10 LIBERO-90 and 4 real	4 LIBERO suites, 10 tasks each	4 LIBERO suites, 10 tasks each
Demonstrations per task	$50$	–	–

Appendix

Table A3: Configurations of the three frozen policies. Execution horizons are reported for the original

1\times

setting. Demonstration counts are provided only for the Flow Matching policies trained in this work; dashes indicate counts not listed for the released checkpoints.

$\rho$	$0.2$	$0.3$	$0.5$	$1$	$1.5$	$2$	$3$	$4$
$f_{\mathrm{s}}$	$0.20$	$0.30$	$0.50$	$1.00$	$1.35$	$1.80$	$2.70$	$3.60$
$f_{\mathrm{c}}$	$0.35$	$0.50$	$0.80$	$1.00$	$1.00$	$1.00$	$1.00$	$1.00$

Appendix

Table A4: Scaling endpoints. One pair per commanded rate, shared by every policy and both simulation and the real robot.

Type	Task	$1\times$	$1.5\times$			$2\times$			$0.5\times$		$0.3\times$
		Ref.	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	Free [-2pt]Speed	w/o [-2pt]CR	Free [-2pt]Speed
Structure constrained	2	$121$	$100$	$98$	$105$	$119$	$93$	$100$	$242$	$194$	$440$	$284$
	32	$204$	$148$	$160$	$178$	$127$	$125$	$152$	$364$	$327$	$580$	$489$
	40	$203$	$158$	$176$	$161$	$161$	$149$	$175$	$360$	$338$	$584$	$560$
	86	$157$	$127$	$116$	$123$	$109$	$105$	$110$	$279$	$255$	$448$	$409$
Bounded receptacle	18	$188$	$138$	$136$	$153$	$125$	$112$	$140$	$336$	$328$	$557$	$497$

Appendix

Table A5: Completion steps under fixed-rate execution on 10 LIBERO-90 tasks with Flow Matching. Each entry reports the mean number of completion steps over that condition’s successful rollouts. Tasks are identified by their LIBERO-90 indices and described in Table A2 . Vanilla denotes vanilla resampling. Success rates and realized rates are reported in Table 2 .

(a) Robomimic: speedup
	$1\times$	$2\times$			$3\times$			$4\times$
Statistic	Ref.	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	vanilla	Free [-2pt]Speed
Overall summary: three tasks
Success (%)	$74.0$	$64.0$	$66.0$	$\mathbf{76.7}$	$32.0$	$35.3$	$\mathbf{60.7}$	$10.0$	$25.3$	$\mathbf{55.3}$
Rate ( $\times$ )	$1.000$	$1.811$	$1.702$	$1.315$	$1.755$	$1.545$	$1.428$	$2.124$	$2.041$	$1.473$

Appendix

Table A6: Fixed-stride execution on Robomimic with frozen Diffusion Policies. Success is aggregated over Lift, Can, and Square using the original behavior-cloning checkpoints, with

50

trials per task and condition. Rate is the equal-task mean of reference/method step ratios on jointly successful trials. Bold marks the highest success rate under each command.

(a) CALVIN: speedup
	$1\times$	$2\times$			$3\times$			$4\times$
Statistic	Ref.	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	vanilla	Free [-2pt]Speed
Overall summary: five-instruction chains, 34 task types
Success (%)	$63.4$	$62.7$	$62.6$	$\mathbf{63.4}$	$59.8$	$55.0$	$\mathbf{64.7}$	$54.5$	$57.1$	$\mathbf{64.9}$
Rate ( $\times$ )	$1.000$	$1.712$	$1.668$	$1.388$	$2.038$	$2.076$	$1.778$	$2.276$	$2.474$	$2.009$

Appendix

Table A7: Fixed-stride execution on CALVIN with frozen

\pi_{0.5}

(ABC

\rightarrow

D). Each stride setting is evaluated over

1000

trials with randomly sampled seeds. Success requires completing all five instructions in a sequence. Realized rates are computed from paired reference/method step ratios, with equal weighting across the

34

task types. Bold marks the highest success rate under each command.

Endpoint	Policy	Task	Command	Realized rate	Ref. success (%)	FreeSpeed success (%)
Minimum	Flow Matching	LIBERO-90, task 48	$0.3\times$	$0.35\times$	$92$	$100$
	$\pi_{0.5}$	LIBERO-10, task 6	$0.2\times$	$0.22\times$	$92$	$92$
	Fast-WAM	LIBERO-Spatial, task 6	$0.2\times$	$\mathbf{0.22\times}$	$100$	$100$
Maximum	Flow Matching	LIBERO-90, task 18	$2\times$	$1.37\times$	$98$	$98$
	$\pi_{0.5}$	LIBERO-Goal, task 5	$3\times$	$2.17\times$	$100$	$100$
	Fast-WAM	LIBERO-Goal, task 0	$4\times$	$\mathbf{2.53\times}$	$100$	$100$

Appendix

Table A8: Endpoints of the per-task success-preserving rate range. For each policy, we report the lowest and highest realized rates among task-command settings where FreeSpeed’s observed success rate matches or exceeds the same task’s

1\times

reference. Success rates

\%

are based on

50

rollouts per task and condition. Bold indicates the overall endpoints,

0.22\times

and

2.53\times

		Ref.	constant factor $f_{t}$
Task		$1\times$	$1.0$	$1.1$	$1.2$	$1.3$	$1.4$	$1.5$	$1.6$	$1.7$	$1.8$	$1.9$	$2.0$	Free [-2pt]Speed
34	Succ.	$98$	$100$	$100$	$100$	$100$	$96$	$92$	$84$	$70$	$62$	$30$	$40$	$\mathbf{100}$
34	Rate	$1.000$	$1.19$	$1.26$	$1.32$	$1.35$	$1.38$	$1.40$	$1.13$	$0.98$	$1.06$	$0.99$	$1.18$	$1.36$
48	Succ.	$92$	$100$	$100$	$94$	$92$	$92$	$96$	$88$	$88$	$88$	$72$	$46$	$\mathbf{98}$
48	Rate	$1.000$	$1.05$	$1.17$	$1.27$	$1.31$	$1.33$	$1.37$	$1.45$	$1.48$	$1.54$	$1.45$	$1.50$	$1.35$
86	Succ.	$92$	$92$	$94$	$92$	$90$	$88$	$78$	$56$	$56$	$54$	$40$	$40$	$\mathbf{88}$

Appendix

Table A9: Constant-factor versus adaptive rescaling. Results on LIBERO-90 tasks 34, 48, and 86 at the

2\times

command, with

50

rollouts per task and condition. The factor

f

remains constant throughout each rollout and is swept from

1.0

2.0

in increments of

0.1

. The setting

f=2

corresponds to w/o CR. FreeSpeed adapts the rescaling factor at each replan using Equation ( 5 ).

Figure A2: Trade-off between success rate and realized rate under constant-factor rescaling. Results are averaged equally across tasks 34, 48, and 86 at the

2\times

command. The left and right axes show success rate and realized rate, respectively. Horizontal dashed lines indicate the results of FreeSpeed and the

1\times

reference.

Figure A3: Success-weighted realized rate for each task. The joint score

J=S\hat{r}

combines success rate

S\in[0,1]

and realized rate

\hat{r}

. Each panel shows one task at the

2\times

command, with

50

rollouts per condition. The open marker identifies the highest-scoring constant factor within the sweep:

1.3

1.8

, and

1.4

for tasks 34, 48, and 86, respectively.

Figure A4: Fixed-stride evaluations shown as curves. Results from Tables 2 and 3 , plotted against the requested execution rate. The upper row shows mean success rates, and the lower row shows realized execution rates. The curves illustrate how each method’s success rate deviates from the

1\times

reference as the requested rate changes.

	$1\times$	$2\times$			$3\times$			$4\times$			$0.5\times$		$0.3\times$		$0.2\times$
	Ref.	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	vanilla	Free [-2pt]Speed	w/o [-2pt]CR	Free [-2pt]Speed	w/o [-2pt]CR	Free [-2pt]Speed	w/o [-2pt]CR	Free [-2pt]Speed
Task 1: Peach
Succ.	$100$	$70$	$90$	$\mathbf{100}$	$50$	$55$	$\mathbf{100}$	$45$	$0$	$\mathbf{95}$	$70$	$\mathbf{100}$	$65$	$\mathbf{100}$	$20$	$\mathbf{95}$
Steps	$324$	$219$	$233$	$234$	$182$	$211$	$204$	$206$	–	$202$	$545$	$433$	$864$	$665$	$1252$	$806$
Rate	$1.00$	$1.48$	$1.39$	$1.39$	$1.77$	$1.53$	$1.59$	$1.57$	–	$1.60$	$0.59$	$0.75$	$0.38$	$0.49$	$0.26$	$0.40$
Task 2: Tennis

Appendix

Table A10: Full fixed-rate results on four real-robot tasks with Flow Matching. Entries report success rates in %, mean completion steps over each condition’s successful rollouts, and realized rates relative to the

1\times

reference. A dash indicates that no trial succeeded. Success averages include all four tasks; completion-step and rate averages include only tasks with successful trials. Bold indicates the highest success rate under each command. Vanilla denotes vanilla resampling, which coincides with w/o CR under deceleration.

Panel of Figure 26 — Figure A5: Observation distribution shift under the $0.3\times$ command. Panels (a) and (b) repeat the principal-component projection and empirical distance CDF of Figure 6 for deceleration. Distances are averaged over the $k=5$ nearest training vectors in the full conditioning space. Under deceleration, w/o CR coincides with vanilla resampling, so only three execution rules are shown.

Figure A7: Accelerated rollouts of the peach-placing task. One rollout is shown per condition. The reference is sampled every $4.7$ seconds of wall-clock time. Each accelerated row shows four frames under the $2\times$ command and three under the $4\times$ command, sampled across the episode. At $4\times$ , w/o CR releases the peach beside the plate, while vanilla resampling throws it; the final frame of the latter rollout shows the peach in the air.

Figure A8: Decelerated rollouts of the peach-placing task. One rollout is shown per condition, sampled every $5.5$ seconds of wall-clock time. The displayed slowdown commands are $0.3\times$ and $0.2\times$ . Under deceleration, vanilla resampling and w/o CR are equivalent, so only w/o CR is shown alongside FreeSpeed.

Figure A9: Accelerated rollouts of the tennis-ball task. One rollout is shown per condition. The reference is sampled every $8.4$ seconds of wall-clock time. Each accelerated row shows four frames under the $2\times$ command and three under the $4\times$ command, sampled across the episode. At $2\times$ , vanilla resampling completes the task but nearly topples the can. At $4\times$ , w/o CR drops the ball beside the can, while vanilla resampling leaves it behind the can, where it appears to overlap the opening from this camera angle.

Figure A10: Decelerated rollouts of the tennis-ball task. One rollout is shown per condition, sampled every $5.9$ seconds of wall-clock time, with frames after the scene settles omitted. At $0.5\times$ , vanilla resampling carries the ball over the can but releases it onto the table beside it.

Figure A11: Accelerated rollouts of the cube-stacking task. One rollout per condition. The reference is sampled every $4.4$ seconds of wall-clock time, and each accelerated row is sampled at four points under the $2\times$ command and three under the $4\times$ one, spread over the episode.

Figure A12: Decelerated rollouts of the cube-stacking task. One rollout per condition, sampled every $8.0$ seconds of wall-clock time.

Figure A13: Accelerated rollouts of the almond-pouring task. One rollout per condition. The reference is sampled every $6.0$ seconds of wall-clock time, and each accelerated row is sampled at four points under the $2\times$ command and three under the $4\times$ one, spread over the episode.

Figure A14: Decelerated rollouts of the almond-pouring task. One rollout per condition, sampled every $17.2$ seconds of wall-clock time.

Explore similar work

Aug 10, 2026cs.RO

SpeedTuning: Speeding Up Policy Execution with Lightweight Reinforcement Learning

While learned robotic policies hold promise for advancing generalizable manipulation, their practical deployment is often hindered by suboptimal execution speeds. Imitation learning policies are inherently limited by hardware constraints and the speed of the operator during data collection. In addition, there are no established methods for accelerating policies learned via imitation, and the empirical relationship between execution speed and task success remains underexplored. To address these issues, we introduce SpeedTuning, a reinforcement learning framework specifically designed to enhance the speed of manipulation policies. SpeedTuning learns to predict the optimal execution speed for actions, thereby complementing a base policy without necessitating additional data collection. We provide empirical evidence that SpeedTuning achieves substantial improvements in execution speed, exceeding 2.4x speed-up, while preserving an adequate success rate compared to both the original task policy and straightforward speed-up methods such as linear interpolation at a fixed speed. We evaluate our approach across a diverse set of dynamic and precise tasks, including pouring, throwing, and picking, demonstrating its effectiveness and robustness in enhancing real-world robotic manipulation. Videos and code are available at https://daivdyuan.github.io/speed-tuning/

David D. Yuan, Tony Z. Zhao, Kaylee Burns +1

Stanford University

Robotic Manipulation Policies Robotic Manipulation Imitation Smooth Execution

Ask PDF

Jul 1, 2026cs.RO

AutoSpeed: Annotation-Free Stage-Adaptive Motion Speed Learning for Robot Manipulation

Different stages of manipulation tasks exhibit varying levels of difficulty, suggesting stage-dependent motion speeds and temporal prediction horizons. However, existing IL-based visuomotor policies typically imitate the execution speed of expert demonstrations and operate with a fixed temporal prediction horizon, limiting flexibility and overall task throughput. In this paper, we introduce AutoSpeed, a model-agnostic learning framework that enables existing visuomotor policies to predict trajectories with stage-adaptive motion speeds, without requiring speed or stage annotations. We treat future trajectories at different speeds as candidate optimization targets, evaluate each candidate using a composite cost that trades off prediction error against prediction horizon, and optimize the policy toward the minimum-cost candidate. With a fixed-length action sequence, speed modulation adjusts the effective temporal prediction horizon: simple stages are executed faster with a longer prediction horizon, whereas complex stages are executed more slowly with a shorter prediction horizon. Specifically, we implement speed modulation in the frequency domain via the discrete cosine transform (DCT), which enables smooth, non-integer speed scaling and thus preserves motion continuity. Extensive evaluations show that AutoSpeed substantially reduces task execution time while also improving success rates. Under the AutoSpeed framework, the inferred motion speeds exhibit a strong correspondence with task stages.

Qingda Hu, Ziheng Qiu, Jieru Zhao +2

College of Intelligent Robotics and Advanced Manufacturing, Fudan University, Shanghai, China · School of Computing, Shanghai Jiao Tong University, Shanghai, China

Visuomotor Policy Robotic Manipulation Motion Action Sequences Short-Horizon Forecasting

Ask PDF

Sep 24, 2026cs.RO

EgoSpeedUp: Transferring Human Manipulation Tempo to Robot Policies

Robot manipulation policies trained through imitation learning inherit not only the demonstrated behavior but also the conservative execution tempo of robot demonstrations. Existing acceleration approaches can execute faster than the original demonstrations, but determine the appropriate acceleration primarily from robot-side information or a predefined set of tempo factors, leaving open how to obtain a task-appropriate reference for how fast each manipulation phase should progress. We introduce EgoSpeedUp, a framework that uses human manipulation as temporal supervision for robot imitation learning. Our key insight is that human demonstrations naturally reveal task-appropriate, phase-wise manipulation tempo. Given slow robot demonstrations and human demonstrations of the same task, EgoSpeedUp aligns corresponding manipulation phases, estimates their relative execution tempos from multiple human demonstrations, and transfers the resulting phase-wise tempo by retiming the robot demonstrations. The retimed demonstrations are then used for standard behavior cloning, allowing the robot to retain its executable manipulation behavior while learning to perform it at a human-informed tempo. Across two real-world manipulation tasks, EgoSpeedUp improves the task success rate by an average of 25 percentage points (pp) while reducing successful execution time by 36.5%. These results demonstrate that human manipulation tempo provides an effective temporal reference for learning faster and more reliable robot policies.

Hanbit Oh, Yukiyasu Domae, Takuma Yagi

Artificial Intelligence Research Center, National Institute of Advanced Industrial Science and Technology (AIST), Japan

Robot-Free Demonstrations Real-World Manipulation Tasks Robot Systems Decode Speedup Human Agency

Ask PDF