ChunkAct: Action Chunking for Agentic Robotics
ChunkAct cuts robot-agent LLM work by batching routine primitives after a sub-goal is chosen.
The paper argues that agentic robotics wastes LLM calls by reasoning over the full history for every primitive action. ChunkAct moves those repeated steps into action chunks: online, one call starts a chained sequence that runs until failure; offline, memory is compiled into chunks so test-time execution does not need memory lookup. On LIBERO-Pro, the authors report a 5.0-point success gain over the prior agentic robotics state of the art, with 2.9x fewer LLM calls, 4.4x fewer input tokens, and 2.3x fewer output tokens. Source: ArXiv · AI/CL/LG's note.
The paper argues that agentic robotics wastes LLM calls by reasoning over the full history for every primitive action. ChunkAct moves those repeated steps into action chunks: online, one call starts a chained sequence that runs until failure; offline, memory is compiled into chunks so test-time execution does not need memory lookup. On LIBERO-Pro, the authors report a 5.0-point success gain over the prior agentic robotics state of the art, with 2.9x fewer LLM calls, 4.4x fewer input tokens, and 2.3x fewer output tokens. Source: ArXiv · AI/CL/LG's note.
score 5