PhysEvo: Astra Can Act, Let It
PhysEvo lets a frozen Astra model recursively improve robot skills, reaching 84% success on real tasks.
PhysEvo is a physical recursive self-improvement framework around a single frozen Astra model: a task agent runs robot tasks and a meta-agent diagnoses failures, revises tools and skills, and can improve its own diagnostics. Across 42 RoboDojo tasks, retained task-specific deployments score 68.14/100 and 62.00% success, versus 47.17% for RoboDawn's one-shot Astra agent; on eight harder manipulation tasks success is 55.00% versus 1.25%. Continuing skill revision after deploying the simulation-evolved harness on an AgileX PiPER yields 90.60/100 and 84.00% success over 25 trials on five real-world tasks.
- PhysEvo improves a frozen Astra model via a task agent and meta-agent.
- Revisions cover tools, skills, and the meta-agent's own diagnostics.
- No weight updates or separately trained action policy are used.
- Held-out RoboDojo success is 62.00% versus 47.17% for one-shot Astra.
- Real AgileX PiPER trials reach 84.00% success across five tasks.
- Products
- AgileX PiPER
- AI models
- Astra
From the original article
1 min read at huggingface.co
Astra can act, yet reliable manipulation depends on the system through which it observes and controls the world. We introduce PhysEvo, a framework for physical recursive self-improvement (RSI) around a single frozen model. A task agent executes robot tasks; a meta-agent uses the resulting trajectories to diagnose failures, revise tools and skills, and test corrections.
By Wenqing Tian, Zeyu Zhang, Zhaocheng Liu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.