AdaTempo: Learning Shared Relative Tempo from Demonstrations for Faster Robot Manipulation

2026-10-01

AI Summary

Imitation learning policies often inherit redundant, slow temporal dynamics from demonstrations, and variations across task phases render uniform acceleration unreliable. This work proposes a phase-aligned trajectory resampling method that leverages the shared relative rhythmic structure across demonstrations to establish phase correspondences and aggregate consensus, thereby generating continuous acceleration curves. By directly embedding the desired tempo into standard policy training, the approach eliminates the need for runtime selection or online retiming. The method generalizes seamlessly to mainstream architectures such as ACT and Diffusion Policy, achieving up to 3.57× speedup while significantly improving the trade-off between success rate and execution speed.

Abstract

Visuomotor policies trained via imitation learning often inherit the unnecessarily slow timing of teleoperated demonstrations. Yet uniform speedup is unreliable because different phases of a manipulation task tolerate acceleration differently. In this work, we introduce AdaTempo, a self-supervised method that accelerates visuomotor policies by exploiting shared relative-tempo structure in demonstrations. AdaTempo establishes phase correspondence, aggregates the aligned relative tempo into a consensus, and maps it to a continuous speedup profile used to resample demonstrations into accelerated training trajectories. Training standard policies such as ACT or Diffusion Policy on these resampled trajectories directly embeds the desired tempo in the learned behavior, without runtime tempo selection or online retiming. Extensive evaluations show that AdaTempo achieves up to a $3.57\times$ speedup and yields a stronger success--speed trade-off than the original policies and representative acceleration baselines.

Problem

• robot manipulation
• visuomotor policies
• imitation learning
• temporal acceleration
• demonstration speedup

Innovation

• AdaTempo
• Relative Tempo
• Self-supervised Learning
• Visuomotor Policies
• Demonstration Resampling