TBNAI
PhysEvo: Astra Can Act, Let It
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 05/10/2026, 20:00:00 UTC
Briefing created: 05/11/2026, 06:57:33 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 05:18:24 UTC.
Briefings
PhysEvo is a framework for physical recursive self-improvement using a single frozen model, where a task agent executes robot tasks and a meta-agent diagnoses failures to revise tools and skills. This process generates joint-level control without model-weight updates. Across 42 RoboDojo tasks, PhysEvo achieved a five-dimension average score of 68.14/100 with 62.00% success, significantly outperforming RoboDawn's one-shot Astra agent which scored 47.17% success. On eight manipulation tasks, PhysEvo reached 55.00% success compared to 1.25% for direct Astra. Deployment on AgileX PiPER yielded 90.60/100 average scores and 84.00% success across 25 trials on five real-world tasks.
Key points
- PhysEvo uses a task agent to execute robot tasks and a meta-agent to diagnose failures, revise tools, and test corrections based on resulting trajectories.
- PhysEvo achieves 62.00% success on 42 RoboDojo tasks, significantly outperforming RoboDawn's one-shot Astra agent which scored 47.17% success.
Context
PhysEvo represents a new approach to physical recursive self-improvement (RSI) by leveraging a single frozen model to evolve through action and observation. Unlike traditional methods requiring model updates, PhysEvo generates persistent, testable changes to how a frozen model acts and improves, enabling joint-level control without separate training.
Limitations
Evaluation is limited to 42 RoboDojo tasks with held-out-layout metrics; direct Astra performance on eight manipulation tasks shows low success rates (1.25%). Real-world deployment on AgileX PiPER shows high scores but limited to 25 trials on five tasks. No additional background context is supplied beyond the provided evidence.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 05/11/2026, 06:57:33 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiefvubbjw74tp7btdu4jp6t4zs46rnskqjpu32avw34kr4smstedi