A robot agent called ChunkAct averaged 8.3 calls to its language-model planner per LIBERO-Pro simulation episode, compared with 24.2 for Harness VLA, while recording higher task success. The version-one preprint, submitted Oct. 13, reports 83.8% success for ChunkAct versus 78.8% for Harness VLA; it includes no physical-robot evaluation.
Turn repeated procedures into reusable chunks
A robot agent often asks a high-level language model to choose what to do next, then passes the decision to perception and control tools. Repeating that reasoning at every step can mean more planner calls and more text sent to the model. Before a task begins, ChunkAct turns a stored procedure into a reusable sequence of perception and control actions, called an action chunk.
At test time, the planner identifies a target and starts a suitable chunk. The chunk chains perception and control steps through a routine procedure. If a step fails, control returns to the planner, which can recover, retry or choose a fallback. The planner still handles decisions and exceptions; the method aims to avoid calling it for every step of a familiar procedure.
ChunkAct used fewer planner calls and had higher simulated success
The authors compared ChunkAct with Harness VLA on LIBERO-Pro, a simulation benchmark that varies task suites and environmental perturbations. Both systems used GPT-5.6-sol at high reasoning, with the same planner, vision-language-action primitive, perception system and task/global memory. The reported difference was how the systems used procedural memory: ChunkAct compiled it into executable chunks offline, while Harness VLA retrieved it during task execution.
The main comparison covered eight suite-and-perturbation combinations, with 10 tasks per combination and 10 seeds per task, or 800 episodes for each system. ChunkAct succeeded in 83.8% of episodes, 5 percentage points above Harness VLA's 78.8%. It averaged 8.3 planner calls per episode versus 24.2, along with 0.5 million input tokens versus 2.2 million and 3,700 output tokens versus 8,500. These are the paper's benchmark averages, not measurements of deployed service cost.
The paper reports simulation results only
The paper reports simulator experiments only; it does not test ChunkAct on a physical robot. Its results are author-reported in a version-one preprint. The article reports no independent replication. The study therefore shows that compiling repeated procedures reduced planner use in this benchmark, but it does not establish real-world reliability, safety or speed.
Sources
- Full paper: “ChunkAct: Action Chunking for Agentic Robotics” (arXiv, version 1)
- arXiv record and abstract, submitted Oct. 13, 2026
- Illustrative image: Nathan J. Szymanski, Bernardus Rendy, Yuxing Fei, Rishi E. Kumar, Tanjin He, David Milsted, Matthew J. McDermott, Max Gallant, Ekin Dogus Cubuk, Amil Merchant, Haegyeom Kim, Anubhav Jain, Christopher J. Bartel, Kristin Persson, Yan Zeng and Gerbrand Ceder, robot arm R1 in a separate autonomous materials laboratory, Wikimedia Commons, CC BY 4.0. This is not the ChunkAct system or experiment.