Speedup Patch: Learning a Plug-and-Play Policy to Accelerate Embodied Manipulation

Zhichao Wu; Junyin Ye; Zhilong Zhang; Yihao Sun; Haoxin Lin; Jiaheng Luo; Haoxiang Ren; Lei Yuan

cs.ROMar 21, 2026

Speedup Patch: Learning a Plug-and-Play Policy to Accelerate Embodied Manipulation

Authors: Zhichao Wu, Junyin Ye, Zhilong Zhang, Yihao Sun, Haoxin Lin, Jiaheng Luo, Haoxiang Ren, Lei Yuan,

Organizations: National Key Laboratory for Novel Software Technology, Nanjing University, China · School of Artificial Intelligence, Nanjing University, China · Mila-Quebec AI Institute & Université de Montréal, Canada

Abstract

While current embodied policies exhibit remarkable manipulation skills, their execution remains unsatisfactorily slow as they inherit the tardy pacing of human demonstrations. Existing acceleration methods typically require policy retraining or costly online interactions, limiting their scalability for large-scale foundation models. In this paper, we propose Speedup Patch (SuP), a lightweight, policy-agnostic framework that enables plug-and-play acceleration using solely offline data. SuP introduces an external scheduler that adaptively downsamples action chunks provided by embodied policies to eliminate redundancies. Specifically, we formalize the optimization of our scheduler as a Constrained Markov Decision Process (CMDP) aimed at maximizing efficiency without compromising task performance. Since direct success evaluation is infeasible in offline settings, SuP introduces World Model based state deviation as a surrogate metric to enforce safety constraints. By leveraging a learned world model as a virtual evaluator to predict counterfactual trajectories, the scheduler can be optimized via offline reinforcement learning. Empirical results on simulation benchmarks (Libero, Bigym) and real-world tasks validate that SuP achieves an overall 1.8x execution speedup for diverse policies while maintaining their original success rates.

Figures & tables

Figure 1: Plug-and-Play Speedup via Scheduler Policy. The scheduler policy $\pi$ predicts a downsampling rate $k$ to downsample the action chunk from the frozen policy into a shorter chunk for acceleration.

Figure 2: Success rate and Violation count. Each subplot (a–d) illustrates the relationship between the cumulative count of violations ( $h_{\mathcal{E}}=1$ ) and the task success rate across different LIBERO suites. The bars represent the conditional success rate for the subset of trajectories containing at least $x$ violations, with the sample size of each subset annotated above the corresponding bar.

Figure 3: The training process of SuP. Our method is trained purely on offline datasets through three phases: (1) recurrent world model learning; (2) data synthesis; and (3) scheduler optimization via IQL. Through this pipeline, we optimize the scheduler to achieve the maximum possible speedup while preserving comparable performance.

Figure 4: Simulation Tasks. We systematically evaluate SuP across 20 tasks from Bigym and 4 task suites (40 tasks in total) from Libero.

Method		Sandwich Remove	Take Cups	Put Cups	Drawers Close All	(15 more tasks) $\dots$	Cupboards Close All	Average
ACT	-base	0.45, 340.5	0.15, 288.3	0.28, 320.3	1.0, 100.0	$\dots$	1.0, 449.8	0.66, 1.00 $\times$
	-ds2	0.48, 186.3	0.13, 178.0	0.36, 175.9	1.0, 52.0	$\dots$	1.0, 234.0	0.61, 1.65 $\times$
	+ DemoSpeedup	0.56, 171.5	0.10, 183.9	0.33, 169.3	1.0, 54.0	$\dots$	1.0, 202.1	0.61, 2.21 $\times$
	+ SuP(Ours)	0.64 , 155.9	0.20 , 176.6	0.38 , 156.0	1.0, 40.0	$\dots$	1.0, 212.1	0.67 , 2.01 $\times$
DP	-base	0.40, 376.8	0.07, 284.6	0.28 , 307.7	0.66, 118.9	$\dots$	0.90, 544.0	0.51, 1.00 $\times$
	-ds2	0.40, 209.6	0.12, 218.0	0.22, 181.5	0.54, 64.9	$\dots$	0.75, 270.3	0.40, 1.42 $\times$

Table 1: SuP speedup results on Bigym Tasks compared with baselines. Each cell in the table reports two metrics: the success rate followed by the average steps to completion, where only successful trajectories are counted. Higher success rates and lower step counts indicate better performance. Cells highlighted in orange denote the best success rate for each task, while those in gray indicate a performance drop exceeding 5% compared to the best success rate. 4 task results are shown, see all task results in App. C .

Method		Spatial	Long	Goal	Object	Average
$\pi_{0.5}$	-base	0.988 , 105.3	0.924, 267.9	0.980, 113.1	0.982, 138.1	0.969, 1.00 $\times$
	-ds2	0.914, 67.9	0.874, 153.4	0.952, 67.6	0.970, 75.0	0.928, 1.72 $\times$
	+ DemoSpeedup	0.964, 88.1	0.932, 221.4	0.968, 88.7	0.988, 114.1	0.963, 1.22 $\times$
	+ SuP(Ours)	0.972, 70.4	0.940 , 215.2	0.986 , 93.4	0.994 , 83.0	0.973 , 1.35 $\times$
VLA- Adapter	-base	0.922 , 99.6	0.936 , 255.1	0.970 , 107.0	0.942, 136.2	0.942 , 1.00 $\times$
	-ds2	0.802, 57.1	0.834, 147.6	0.930, 57.1	0.882, 76.3	0.862, 1.77 $\times$

Table 2: SuP speedup results on Libero compared with baselines. Each cell in the table reports two metrics: the success rate followed by the average steps to completion, where only successful trajectories are counted. Higher success rates and lower step counts indicate better performance. Cells highlighted in orange denote the best success rate for each task, while those in gray indicate a performance drop exceeding 1% compared to the best success rate.

Method	Training Params	Training Time	Inference Overhead
DemoSpeedup	4B	20h	-
SuP (Ours)	5.12M	2h	1ms (2%)

Table 3: Computational Efficiency of SuP in Libero.

Figure 5: Real-world Tasks Illustration. We illustrate the procedure of 3 real-world tasks: (a) Arange Table (b) Fold Towel (c) Stack Plates .

Method		Arrange Table	Fold Towel	Stack Plates	Average
$\pi_{0.5}$	-base	11/30, 537.8	15/30, 519.5	27/30, 221.5	0.589, 1.00 $\times$
	+ ds2	10/30, 291.9	14/30, 326.3	27/30, 177.1	0.567, 1.61 $\times$
	+ ds3	3/30, 247.2	8/30, 187.7	21/30, 148.4	0.356, 2.19 $\times$
	+ DemoSpeedup	12/30, 267.4	14/30, 223.2	28/30 , 124.4	0.600, 2.07 $\times$
	+ SuP(Ours)	13/30 , 258.5	16/30 , 192.5	26/30, 138.5	0.611 , 2.17 $\times$

Table 4: SuP speedup results on Real-world Tasks compared with baselines.

Figure 6: Case study. We visualize the adaptive downsampling strategy during a Fold Towel task. The plot tracks the predicted downsampling rate over the episode timesteps. Shaded regions annotate distinct task phases.

Figure 7: Comparison between ADM-based and MLP-based world model. (a) Visualization of task success rates against predicted violation counts in the Spatial suite. (b) Spearman correlation scores across four LIBERO suites.

Ablation	Sandwich Remove	Put Cup	Long
$\pi_{base}$	0.45, 340.5	0.28, 320.3	0.924, 267.9
SuP-0.01	0.59, 187.5	0.35, 159.8	0.940 , 215.2
SuP-0.015	0.64 , 155.9	0.38 , 156.0	0.922, 191.8
SuP-0.02	0.46, 157.6	0.21, 143.3	0.904, 168.4
MPC-0.01	0.54, 181.4	0.32, 154.8	0.936, 225.2
MPC-0.015	0.59, 173.3	0.3, 146.1	0.924, 205.6

Table 5: Ablation study on SuP. We compare the performance impact of the IQL module and different deviation thresholds

\epsilon

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Method		Sandwich Remove	Take Cups	Put Cups	Dishwasher Open Trays	Move Plate
ACT	-base	(0.45, 340.5)	(0.15, 288.3)	(0.28, 320.3)	( 1.0, 275.0)	( 0.58 , 194.8)
	-ds2	(0.48, 186.3)	(0.13, 178.0)	(0.36, 175.9)	( 1.0, 169.0)	(0.50, 155.0)
	+ DemoSpeedup	(0.56, 171.5)	(0.10, 183.9)	(0.33, 169.3)	(1.0, 156.0)	(0.28, 71.7 )
	+ SuP(Ours)	( 0.64 , 155.9 )	( 0.20 , 176.6 )	( 0.38 , 156.0 )	(1.0, 131.0 )	(0.49, 167.2)
DP	-base	(0.40, 376.8)	(0.07, 284.6)	( 0.28 , 307.7)	(0.57, 351.2)	(0.33, 261.1)
	-ds2	(0.40, 209.6)	(0.12, 218.0)	(0.22, 181.5)	(0.48, 310.2)	(0.36, 203.3)

Appendix

Table 6: Performance comparison of different methods in all Bigym tasks. Best success rates and shortest lengths are bolded . Method with best success rate per task is highlighted .

Table 7: Hyperparameters for ACT and DP in Bigym.

Figure 8: ARX5 illustration. (a) The master-follow design for data collection (b) The actual scene of our real-world experiment.

Method	spatial	long	goal	object	Average
$\pi_{0.5}$ (Original)	0.988, 105.3	0.924, 267.9	0.980, 113.1	0.982, 138.1	0.969, 1.00 $\times$
No Grip Fix (-ds2)	0.708, 77.0	0.818, 167.4	0.888, 66.2	0.954, 84.4	0.842, 1.58 $\times$
With Grip Fix (-ds2)	0.914, 67.9	0.874, 153.4	0.952, 67.6	0.970, 75.0	0.928, 1.72 $\times$

Appendix

Table 8: Ablation study of gripper action compensation on Libero benchmarks. We compare the original policy (

pi_{0.5}

) with downsampled versions (

N=2

) before and after applying the gripper fix.

Figure 9: Network architecture of Recurrent World Model and Scheduler Policy.

Hyperparameter	BiGym	Libero	Real-world
Learning rate	$3\times 10^{-4}$	$1\times 10^{-4}$	$1\times 10^{-4}$
Batch size	512	512	512
GRU hidden dimension	256	256	256
GRU layers	3	3	3
Chunk length	24	10	20
$k_{min}$	2	1	2

Appendix

Table 9: Hyperparameter configurations of SuP in different experiment settings.

Robot Policy Learning

Robot policy learning develops controllers that map sensory observations, task goals, and interaction history to actions for embodied robots.

See all 372 papers

SpeedTuning: Speeding Up Policy Execution with Lightweight Reinforcement Learning
Aug 10, 2026David D. Yuan, Tony Z. Zhao, Kaylee Burns +1Robot Policy Learning Robot Skill Learning
Stanford University
FreeSpeed: Training-Free Speed Control for Generative Robot Policies
Oct 5, 2026Yuxuan Hu, Shilin Shan, Qiheng Wang +4Robot Policy Adaptation Robotic Manipulation
MARS Lab, Nanyang Technological University · ROKAE Robotics
EgoSpeedUp: Transferring Human Manipulation Tempo to Robot Policies
Sep 24, 2026Hanbit Oh, Yukiyasu Domae, Takuma YagiRobot Policy Learning Robotic Manipulation
Artificial Intelligence Research Center, National Institute of Advanced Industrial Science and Technology (AIST), Japan

Aug 10, 2026J/K move · Enter open · S save

SpeedTuning: Speeding Up Policy Execution with Lightweight Reinforcement Learning

David D. Yuan, Tony Z. Zhao, Kaylee Burns, Chelsea Finn

Stanford University

Robot Policy Learning Robot Skill Learning Policy Optimization Robotic RL Robotic Manipulation

Open paper AskPDF arXiv

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/

Abstract

Robot Policy Learning

Explore similar work

SpeedTuning: Speeding Up Policy Execution with Lightweight Reinforcement Learning

FreeSpeed: Training-Free Speed Control for Generative Robot Policies

EgoSpeedUp: Transferring Human Manipulation Tempo to Robot Policies