Reproduce “Dynamic Execution Horizon Prediction for Chunk-based Robot Policies” 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
Yuchi Zhao, Miroslav Bogdanovic, Arjun Sohal, Liyu Tao, Kourosh Darvish, Alán Aspuru-Guzik, Florian Shkurti, Animesh Garg
Researchers from the University of Toronto and Vector Institute developed Dynamic Execution Horizon Prediction (DEHP), a module that dynamically adjusts a robot's replanning frequency, which consistently improved success rates for chunk-based policies across diverse manipulation tasks and proved robust to varying action noise.
Key insights
- Optimizing the chunk-level discounted objective is theoretically equivalent to optimizing the original discounted return of the base MDP, enabling the use of a standard state-value critic for on-policy learning.
- DEHP learns interpretable, phase-dependent execution schedules, selecting shorter horizons for precision-critical phases (e.g., tight insertions) and longer horizons for less critical free-space movements.
- The modular design allows DEHP to act as an adaptive control layer over existing black-box chunk policies, isolating the horizon adaptation from the action generation mechanism and enabling broad compatibility.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.