Robot learning framework improves long task success and speed
RACaP: Agentic Reasoning, Acting, and Coding as Policies for Evolvable Robot Learning
Robotics
Summary
Robots often struggle with learning new tasks quickly and handling long or complex sequences of actions. The authors developed a method called RACaP that separates reusable robot skills from the decision-making process, letting robots learn and adapt more efficiently. This approach uses a loop of reasoning and acting with fixed skill APIs that can improve over time without changing the original code. Their experiments show this method works better and faster than previous approaches on challenging robot task benchmarks.
What this means in practice
- •For robotics engineers: Develop robots that perform long, complex tasks more reliably by using RACaP’s reusable skill APIs combined with adaptive reasoning loops.
- •For embedded system developers: Deploy faster robot decision-making components by distilling large language model policies into smaller, efficient models using rejection-sampled fine-tuning.
- •For automated manufacturing integrators: Integrate evolvable robot control systems that support task transfer and recovery without rewriting base code, improving factory automation flexibility.$Commercial implications: Enables selling adaptive robot software platforms that reduce downtime and reprogramming costs for automated factories.
Authors
Zexi Li, Yehang Zhang, Haojian Huang, Bohan Zhou, Wenqian Li, Chenxu Wang, Yifan Chang, Yangkai Wei, Tianyi Zhang, Ying-Cong Chen, Kaiwen Zhou, Yinchuan Li, James Cheng
Abstract
General-purpose robot agents must learn from experience, transfer to new tasks, and act efficiently. Code as Policies (CaP) methods generate and repair programs at runtime, incurring latency and entangling reusable mechanisms with task-specific decisions. We introduce RACaP, an agentic framework that moves coding to evolution and uses a Reasoning-and-Acting (ReAct) loop to call frozen, typed Policy APIs at deployment. A two-phase strategy combines capability curriculum learning with autonomous self-evolution to improve the APIs, the ReAct harness, and experience memory. The APIs encode reusable physical mechanisms while exposing arguments for runtime adaptation. ReAct combines task-specific working memory, long-term experience memory, and visual feedback to select actions, verify outcomes, and recover from failures without modifying source code. RACaP achieves 54.4% success on LIBERO-90, 45.0% on zero-shot LIBERO-PRO, and 46.0% on LIBERO-Long, compared with at most 4.0% for CaP baselines on long-horizon tasks. On LIBERO-PRO, it achieves 2.5 times the success rate of CaP baselines and a 1.9-fold speedup in median policy time. For efficient on-robot deployment, rejection-sampled fine-tuning distills GPT-5.6 ReAct decisions into Qwen3-VL-8B-Instruct, yielding a 13.2-fold per-decision inference speedup and reducing repeated physical calls from 16 to 4. These results show that separating reusable code from runtime decisions supports continued evolution, effective transfer, and efficient long-horizon control.