Paper Overview
Field: NLP / Robotics Authors: Bingxin Xu, Wenrui Bao, Xinxin Liu, Jiawei Zhou, Yuzhang Shang, Emilio Ferrara Published: 2026-10-13 arXiv: 2610.18874
Summary
Agentic robotics places a large language model (LLM) agent in the robot control loop, where it composes perception, motion, and vision-language-action (VLA) tools under the guidance of a memory, enabling generalization to new tasks without retraining. This generality comes at a steep efficiency cost: every primitive requires a round of LLM reasoning over the full interaction history, so the cost of an episode grows linearly with the number of primitives.
The authors find this reasoning is largely redundant — once a sub-goal is chosen, its primitives follow a routine already prescribed by memory. ChunkAct introduces action chunking to agentic robotics:
- Online: a single LLM call launches a chunk of chained primitives, which executes autonomously and returns control only upon failure.
- Offline: ChunkAct compiles memory into these chunks, so test-time execution requires no memory lookups, and the planner is reserved only for decisions that actually need reasoning.
- +5.0 percentage points in task success rate
- 2.9x fewer LLM calls
- 4.4x fewer input tokens
- 2.3x fewer output tokens
As a result, LLM computation scales with the number of decisions rather than the number of primitives.
Results
On the LIBERO-Pro benchmark, ChunkAct outperforms state-of-the-art agentic robotics methods:
*Auto-collected on 2026-10-15.*