OpenClaw-RL: Continuous RL Optimization from Live Conversations

Source: https://github.com/Gen-Verse/OpenClaw-RL Author: Gen-Verse Date: 2025-2026

Summary

RL framework that wraps a self-hosted model as an OpenAI-compatible API, intercepts live multi-turn conversations, and continuously optimizes the policy in the background. Fully asynchronous four-component architecture: agent serving, rollout collection, reward evaluation, policy training. All models and data remain on user infrastructure.

Key Claims

  • Self-hosted, private — zero API dependency, all data on user infrastructure
  • OpenAI-compatible API wrapper — drop-in replacement
  • Fully asynchronous: four independent components don’t block each other
  • Three optimization methods: Binary RL, On-Policy Distillation (OPD), combined
  • Results: “significant improvement” with just 36 problem-solving + 24 grading interactions
  • No manual labeling — conversation feedback converted automatically to training signals
  • Scales from personal agents to: terminal agents, GUI agents, software engineering agents

Connection to Other Sources

Practical implementation of RL from User Conversations and Cursor Real-Time RL. Self-hosted version of what labs do in production.

Concepts