ChunkAct
A method that introduces the idea of action chunking into the planning layer of agentic robots. By executing chained primitive action chunks online and compiling memory modules offline, it reduces the number of LLM calls from one per primitive to one per decision, thereby cutting redundant reasoning and token consumption.
Timeline (last 90 days)
ChunkAct improves task success rate on the LIBERO-Pro benchmark by 5.0 percentage points over the current state-of-the-art method in the agentic robotics field
ChunkAct reduces the number of LLM calls by 2.9x on LIBERO-Pro
ChunkAct reduces input tokens by 4.4x and output tokens by 2.3x on LIBERO-Pro
The core method of ChunkAct is to introduce action chunking into agentic robotics, where a single LLM call at online execution time initiates an entire chunk of chained primitives, returning control to the planner only upon failure
In the offline stage, ChunkAct compiles the memory module into action chunks; in the testing stage, there is no need to query memory, and the amount of computation is tied to the number of decisions rather than the number of primitives
The researchers of ChunkAct found that a large portion of LLM reasoning in agentic robotics is redundant, because once a subgoal is selected, the sequence of primitives required for execution follows the routine already specified in the memory module, with no need to re-reason at each step
ChunkAct elevates action chunking from low-level continuous control to the agentic planning level; the object of chunking is the sequence of primitive calls, and the chunk boundaries are determined by decision points rather than fixed time windows
In ChunkAct, the online execution path of the agentic robot is shorter and more deterministic due to offline compilation of action chunks, reducing the propagation of uncertainty caused by LLM reasoning
Reducing LLM calls usually sacrifices task completion quality, but ChunkAct, by eliminating redundant reasoning, enables the planner to focus more on the steps that truly require judgment, thereby improving both success rate and efficiency
All Facts (9)
ChunkAct improves task success rate on the LIBERO-Pro benchmark by 5.0 percentage points over the current state-of-the-art method in the agentic robotics field
50%UnverifiedChunkAct reduces the number of LLM calls by 2.9x on LIBERO-Pro
50%UnverifiedChunkAct reduces input tokens by 4.4x and output tokens by 2.3x on LIBERO-Pro
50%UnverifiedThe core method of ChunkAct is to introduce action chunking into agentic robotics, where a single LLM call at online execution time initiates an entire chunk of chained primitives, returning control to the planner only upon failure
50%UnverifiedIn the offline stage, ChunkAct compiles the memory module into action chunks; in the testing stage, there is no need to query memory, and the amount of computation is tied to the number of decisions rather than the number of primitives
50%UnverifiedThe researchers of ChunkAct found that a large portion of LLM reasoning in agentic robotics is redundant, because once a subgoal is selected, the sequence of primitives required for execution follows the routine already specified in the memory module, with no need to re-reason at each step
50%UnverifiedChunkAct elevates action chunking from low-level continuous control to the agentic planning level; the object of chunking is the sequence of primitive calls, and the chunk boundaries are determined by decision points rather than fixed time windows
50%UnverifiedIn ChunkAct, the online execution path of the agentic robot is shorter and more deterministic due to offline compilation of action chunks, reducing the propagation of uncertainty caused by LLM reasoning
50%UnverifiedReducing LLM calls usually sacrifices task completion quality, but ChunkAct, by eliminating redundant reasoning, enables the planner to focus more on the steps that truly require judgment, thereby improving both success rate and efficiency
50%