ChunkAct: Action Chunking for Agentic Robotics
- Published
- Source
- arXiv
- Paper number
- 1348
- Field
- Robotics
- arXiv ID
- 2610.18874
Key points
- Prior agentic robotics runs one round of LLM reasoning per primitive, so a single pick-and-place task requires 24.2 LLM calls and 2.2M input tokens on average.
- An analysis of the planner's tool calls shows that only 20.5% invoke an action primitive, while memory reads and perception account for 31.7% and 44.3%, respectively.
- ChunkAct compiles memory into executable chunks per task type in advance, and at test time a single call launches a whole chunk that returns control only upon failure.
- On the LIBERO-Pro benchmark it attains an overall success rate of 83.8%, the best result, while requiring 2.9× fewer LLM calls, 4.4× fewer input tokens, and 2.3× fewer output tokens.
- Running the chunks under a scripted driver with no LLM call still reaches 57.5% success, showing that the planner's contribution concentrates on recovery from failure.
- Applied unchanged to LIBERO-Plus with seven perturbation axes unseen during memory construction, it achieves an overall success rate of 83.9%, confirming zero-shot transfer.
Paper links
External research summaries. These are not HDATF publications or measured product results.