Areal: RL Infrastructure for OpenClaw

​Welcome to NICE Talk 154!

​youtube livestream: https://www.youtube.com/watch?v=C-1po0fKbNA

​Talk

​Reinforcement learning (RL) has been one of the key research focuses of the LLM community over the past year, and Agentic RL in 2026 has undoubtedly ignited a competitive race among major model vendors.

​This talk starts from the RL infrastructure perspective, and breaks down by source code analysis the online RL training features introduced in the open-source project AReaL 1.0: users only need to plug in AReaL's base URL and API key to integrate reinforcement learning training with zero code changes.

​This approach supports training and self-evolution for any agent application, including the recently trending OpenClaw. In addition, the talk will explore the challenges and opportunities facing RL infrastructure in current application scenarios, and share AReaL's architectural evolution direction, inspired by the OpenClaw project, amid the current agent wave.

​Speaker:
Wei Fu, Ph.D. student at Tsinghua University

Host:
Wenyue Hua, senior researcher at Microsoft Research