OpenClaw-RL: Train any agent simply by talking

🦞 Just talk to your agent — it learns and EVOLVES 🧬.

Let Skills Evolve Collectively with Agentic Evolver

[CoRL 2023] This repository contains data generation and training code for Scaling Up & Distilling Down

Self-evolving runtime infrastructure for Physical AI and embodied agents. Ground AI agents into robot bodies with e-URDF, sandbox safety, capability routing, praxis capture, physical memory, runtime intervention, and skill evolution.

An agent-first mobile runtime built on OpenClaw: control apps, learn reusable skills, and connect phones.

SRA-Bench and SR-Agents: a benchmark and toolkit for skill-retrieval-augmented LLM agents.

玄霖超脑 · 无量网络 v4 重构版— 66 模块6通道 AI 认知神经系统 · 装一次，你所有的 AI 工具从此共享一个永远不失忆的大脑。跨会话记忆 · 多 Agent 互通 · 自动代码审查 · 零配置

[ICLR 2023] Choreographer: a world-model-based agent that discovers and learns unsupervised skills in latent imagination, and it's able to efficiently coordinate and adapt the skills to solve downstream tasks.

Agent Skill Induction: "Inducing Programmatic Skills for Agentic Tasks"

[ICML 2023] Pre-train world model-based agents with different unsupervised strategies, fine-tune the agent's components selectively, and use planning (Dyna-MPC) during fine-tuning.

PyTorch implementation for "Discovery of Incremental Skills" (DISk) algorithm from ICLR 2022 paper "One After Another: Learning Incremental Skills for a Changing World"

Winning code for the Object Manipulation MyoChallenge Track at NeurIPS 2023

Official implementation of SPECI: Skill Prompts based Hierarchical Continual Imitation Learning for Robot Manipulation [IEEE TCDS]

[Neuron 2024] Analysis of Baoding ball policy from MyoChallenge at NeurIPS 2022

A self-extending AI agent built with Google ADK & Gemini 2.0 Flash that dynamically creates, stores, and reuses its own skills at runtime — growing smarter with every task.

Flow-driven recursive skill evolution for agentic LLM orchestration.

Solving the amnesiac problem for LLM agents. Research series on agents that compound knowledge across sessions — first measurement: +4.6 pp accuracy lift on Terminal-Bench 2.1 with an open-weight executor and a single failure-derived skill file.

Android app for skill learning and management built with Java & Firebase.

Self-evolving skill engine for OpenClaw agents. Observes tool usage, crystallizes patterns into auditable SKILL.md files through a dual-model LLM pipeline, and continuously validates with 23 adversarial mutations. Research-grounded. Human-approved. Nothing auto-deploys.