# SkillClaw

- **Event:** [Nebius.Build SF](https://cerebralvalley.ai/e/nebius-build-sf)
- **When:** Sun, Mar 15 at 9:00 AM – 10:00 PM (PDT)
- **Where:** Shack15, San Francisco, CA
- **Team:** [Lily Zhang](https://cerebralvalley.ai/u/lilyzhng), [Yifei Shao](https://cerebralvalley.ai/u/roboclaw)
- **GitHub:** https://github.com/lilyzhng/SkillClaw
- **Demo video:** https://youtu.be/ZIxJmRpsZzI?si=Ne5k6klZS5Po86jS
- **Gallery:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery/5

Traditional robot learning is slow: collect demonstrations, train a policy, deploy, repeat from scratch for every new task. Knowledge never compounds. Every new task is day zero.

SkillClaw inverts this. Instead of training policies, we let multiple AI agents figure out robot manipulation by themselves. Give them a motion planner, a simulation and error logs in text and image. They try, they fail, they learn from the failure, they try again. Eventually they figure it out. And when they do, that skill gets saved so every future agent can build on it.

Multiple agents work on different tasks in parallel. Each has its own private workspace. When one solves a task, the skill gets packaged into a PR with a demo video, reviewed by an oversight agent, and promoted to a shared brain that all agents can read from. The 7th task is easier than the 1st because the agent already knows how to pick, push, and grasp.

So far: 8 skills learned across 14 robotics benchmark tasks, from simple cube pushing to sub-millimeter peg insertion. Every task any agent solves makes every future agent better. The skill library is the flywheel.

---

Markdown version of https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery/5. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt