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

Qingda Hu^1,* Ziheng Qiu^1,* Jieru Zhao² Zhongxue Gan¹ Wenchao Ding^1,†

¹Fudan University ²Shanghai Jiao Tong University

^*Equal Contribution ^†Corresponding Author

Email: tinda24@163.com, zhqiu25@m.fudan.edu.cn

European Conference on Computer Vision, 2026

AutoSpeed stage-aware motion speed adaptation overview

Motion speeds in expert demonstrations can be suboptimal. AutoSpeed trains policies to execute easy stages with longer effective prediction horizons and higher speeds, while preserving slower, shorter-horizon behavior for fine-grained manipulation stages.

Abstract

Different stages of manipulation tasks exhibit varying levels of difficulty, suggesting stage-dependent motion speeds and temporal prediction horizons. Existing imitation-learning visuomotor policies usually imitate the execution speed of expert demonstrations and predict over a fixed temporal horizon, which limits flexibility and task throughput. AutoSpeed is a model-agnostic learning framework that enables existing visuomotor policies to predict trajectories with stage-adaptive motion speeds without requiring speed or stage annotations. It treats future trajectories at different speeds as candidate supervision targets, evaluates each candidate with a composite cost that balances prediction error and prediction horizon, and optimizes the policy toward the minimum-cost candidate. DCT-based frequency-domain speed modulation supports smooth non-integer speed scaling while preserving motion continuity. Across simulation and real-world evaluations, AutoSpeed reduces task execution time while improving success rates, and the inferred motion speeds correspond closely to task stages.

Highlights

Annotation-Free Predict trajectories with stage-adaptive motion speeds, without requiring speed or stage annotations.

Stage-Adaptive Simple stages are executed faster with a longer prediction horizon, whereas complex stages are executed more slowly with a shorter horizon.

Model-Agnostic Model-agnostic learning framework that applies to non-generative and generative visuomotor policies.

1.7x Speedup Significantly reduces execution time, delivering an average speedup of approximately 1.78× across the diverse real-world task suite.

Method

AutoSpeed cost-aware multi-target selective optimization

Candidate Speed Targets

AutoSpeed builds a set of future action chunks at different speed ratios. Each candidate keeps the model output length fixed while changing the effective prediction horizon.

Cost-Aware Selection

A composite cost that trades off prediction error against temporal prediction horizon. The policy is optimized toward the minimum-cost candidate.

DCT Motion Transform

Speed modulation is implemented in the frequency domain with the discrete cosine transform, enabling smooth acceleration and deceleration with non-integer ratios.

Ratio Head and NTA

A lightweight ratio head predicts the speed ratio from observation, and Nonlinear Temporal Aggregation outputs actions by combining the speed for smooth deployment.

Optimization View

AutoSpeed formulates training as cost-aware Multi-Target Selective Optimization. Candidate future trajectories are generated by retiming the same action chunk with different speed ratios. The model then optimizes toward the candidate with the minimum composite cost, allowing speed preference to emerge from ordinary demonstration data.

Intuitively, in easier stages, future actions remain predictable over longer horizons, permitting higher speed ratios; in precision-critical stages, long-horizon prediction becomes less predictable, yielding lower speed ratios.

Simulation Benchmarks

AutoSpeed is evaluated on 62 simulation tasks from ALOHA Simulation, LIBERO-10, and Meta-World. The experiments cover both single-task and multi-task learning, including non-generative ACT and generative BAKU action heads.

Simulation tasks used in AutoSpeed evaluation

Table. Quantitative results across multiple tasks on the ALOHA benchmark. † denotes evaluated with a high-gain controller.

Method	Transfer Cube		Insertion
Method	SR (↑)	Len (↓)	SR (↑)	Len (↓)
ACT	72%	272	22%	353
ACT(AutoSpeed)	64%	160	24%	300
ACT(AutoSpeed^†)	78%	165	24%	296

Table. Quantitative results on multi-task learning benchmarks. Success rate (SR, ↑) and episode length (Len, ↓).

Method	LIBERO-10		Meta-World
Method	SR (↑)	Len (↓)	SR (↑)	Len (↓)
BAKU-DiT	50%	261	43%	87
BAKU-Flow	47%	287	57%	82
BAKU-DiT (AutoSpeed)	49%	259	62%	70
BAKU-Flow (AutoSpeed)	52%	171	65%	71

Real-World Results

Real robot experiments are conducted on an Agilex Piper bimanual platform with three RGB cameras and robot proprioceptive state. Across four tabletop tasks, AutoSpeed with Nonlinear Temporal Aggregation improves success rates and reduces completion time.

Real-world AutoSpeed task stages and speed curves

Table. Real-world evaluation. Success rate (SR, ↑) and execution time in seconds (Time, ↓).

Method	Transfer the Banana		Place the Toy
Method	SR (↑)	Time (↓)	SR (↑)	Time (↓)
BAKU-Flow+TA	77%	20.80s	33%	18.13s
BAKU-Flow (AutoSpeed)+NTA	83%	11.24s	43%	10.13s
Method	Stacking Two Cubes		Store Block in Drawer
Method	SR (↑)	Time (↓)	SR (↑)	Time (↓)
BAKU-Flow+TA	60%	50.11s	77%	38.67s
BAKU-Flow (AutoSpeed)+NTA	63%	28.52s	80%	23.01s

Stage-Aware Speed Curves

Predicted speed curve for a real-world banana transfer task — Real-world speed-ratio trajectory.

Ablation on AutoSpeed speed range bounds — Speed-range ablation on ALOHA Transfer Cube.

The learned speed curves show consistent phase-aware patterns. Policies accelerate during easier free-space motion and decelerate around complex interaction stages, even when trained with different predefined speed ranges.

More Results

Whac-A-Mole Real-World Experiment

AutoSpeed trained on the Whac-A-Mole game adaptively predicts higher motion speeds while the robot arm moves between targets and lowers the speed during pressing, improving precision in interactions.

BibTeX

@inproceedings{autospeed2026,
  title={AutoSpeed: Annotation-Free Stage-Adaptive Motion Speed Learning for Robot Manipulation},
  author={Hu, Qingda and Qiu, Ziheng and Zhao, Jieru and Gan, Zhongxue and Ding, Wenchao},
  journal={arXiv preprint arXiv:2607.01051},
  year={2026}
}