Reproduce “Learning High-Frequency Continuous Action Chunks in Latent Space” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Kunyun Wang, Yuhang Zheng, Yupeng Zheng, Jieru Zhao, Wenchao Ding

The authors introduce a latent-action policy that predicts motion chunks, decodes them into 60-Hz commands, and uses Reuse-then-Refine (RTR) to smooth asynchronous handoffs. In synchronous OpenVLA-OFT whiteboard writing, latent training raised success from 74% to 100%, reduced jerk from 5.238 to 0.558, and exceed count from 50.5 to 2.2 in 50 trials. Tests cover 60-Hz tasks; higher frequencies and action-space-trained policies remain untested.

Key insights

  • Learning high-frequency actions in a continuous latent space simplifies the policy approximation problem, leading to inherently smoother and more precise trajectories compared to direct action-space learning.
  • A training-free refinement mechanism, Reuse-then-Refine (RTR), can effectively maintain chunk-level continuity under asynchronous inference, mitigating common issues like stalls and jerky movements.
  • The co-design of a robust action representation (latent space via VAE) and an efficient execution strategy (RTR for continuity) is crucial for achieving high-quality, real-time robotic control in complex tasks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.