OpenClaw-RL: The Self-Evolving Agent (NICE No.158)

​Welcome to NICE TALK 158! | OpenClaw-RL: The Self-Evolving Agent

​

​What if your model could learn and evolve from every interaction after deployment?

​

​This talk introduces OpenClaw-RL, a novel reinforcement learning framework that enables agents to self-improve autonomously during real-world use. OpenClaw-RL functions as an RL server—users simply deploy their personal models on it, and the models automatically and continuously optimize throughout usage. We propose an optimization method that combines the strengths of GRPO and On-policy Distillation, transforming the entire history of model-user-environment interactions into RL training signals. The result: personal agents that don't just stay static, but grow more capable and adaptive the more they are used. We also design interesting experiments to validate the framework's efficient optimization capability for personal agents.

​

​Resources:

• ​Paper: https://arxiv.org/abs/2603.10165

• ​Code: https://github.com/Gen-Verse/OpenClaw-RL

​

​Speaker: Yinjie Wang

​Yinjie Wang is a second-year PhD student at the University of Chicago and a Research Intern at the Princeton AI Lab. Previously, he graduated from the Special Class for the Gifted Young (SCGY) at the University of Science and Technology of China. His research focuses on LLMs, autonomous agents, and reinforcement learning. He is the creator of several open-source RL frameworks including OpenClaw-RL, RLAnything, CURE, and dLLM-RL. His work has been published at NeurIPS, ICLR, and other top-tier venues, receiving a Spotlight award at NeurIPS 2025.

​Personal website: https://yinjjiew.github.io/

​

​Host: Wenyue Hua

​Senior Researcher at Microsoft Research.