Robots that think and act using advanced language-based AI often face a big challenge: every tiny step they take requires the AI to carefully analyze the entire task history, making the process slow and resource-heavy. A new research paper introduces a clever solution called ChunkAct, which groups sequences of robot actions into “chunks,” allowing robots to perform multiple steps autonomously before consulting the AI again. This approach promises to make robot decision-making faster and more efficient without sacrificing accuracy, potentially advancing how robots learn and adapt to new tasks.
Key Takeaways
- ChunkAct reduces the number of AI reasoning calls by nearly three times, making robot control more efficient.
- By grouping actions into chunks, robots can execute multiple steps autonomously, only asking the AI for help when necessary.
- ChunkAct improves task success rates by 5 percentage points compared to previous methods.
- The approach lowers the amount of data (input and output tokens) the AI needs to process, cutting computational costs.
The research tackles a fundamental issue in “agentic robotics,” where a large language model (LLM)—a type of AI trained on vast amounts of text—is integrated into robot control. These models help robots understand and plan complex sequences by combining perception (what the robot senses), motion (how it moves), and vision-language-action tools (how it interprets visual info and language commands). While powerful, this setup demands that the AI repeatedly reason over the entire history of the robot’s interactions for every small action it takes, which quickly becomes inefficient.
ChunkAct introduces the idea of “action chunking.” Instead of the AI reviewing and deciding every single primitive action—basic steps like moving a robotic arm or grasping an object—the system groups these into chunks, or chains of actions, that the robot can carry out on its own once a sub-goal is set. The AI intervenes only when the robot encounters a problem or failure, significantly reducing how often the AI needs to process information.
To achieve this, ChunkAct works in two phases. Offline, it compiles the robot’s memory into these action chunks, essentially learning routine sequences that can be reliably executed without constant AI oversight. Online, during actual task execution, a single call to the language model launches a whole chunk of actions that run autonomously. This means the AI’s computational workload scales with the number of decisions or sub-goals, not with the number of individual actions, making the system more scalable and faster.
Testing ChunkAct on the LIBERO-Pro benchmark—a standard set of tasks designed to evaluate robotic agents—showed clear improvements. The method not only increased the success rate by 5 percentage points over the previous state-of-the-art agentic robotics approach but also drastically reduced the number of AI calls and the size of data processed. Specifically, it required 2.9 times fewer calls to the language model, 4.4 times fewer input tokens, and 2.3 times fewer output tokens, highlighting significant efficiency gains.
These improvements could have meaningful implications for the future of robotics, especially in areas requiring flexible, adaptive robots that can learn new tasks without retraining. By cutting down on the AI’s processing demands, ChunkAct opens the door to more practical deployment of language-driven robotic systems in real-world environments, from manufacturing to home assistance. Moving forward, researchers may explore extending action chunking to even more complex tasks and integrating it with other AI advances to further enhance robot autonomy and efficiency.
Based on research published on arXiv by Bingxin Xu, Wenrui Bao, Xinxin Liu et al..