TinyAstral distills harness improvements to 4B robot agents
Original title: TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents
TinyAstral transfers gains from large teacher harnesses to compact students reaching near teacher performance on manipulation tasks.
TinyAstral improves a teacher's harness via recursive self-improvement where frontier agents propose edits accepted only after statistical tests. It then distills to a 4B student in the same harness. On RoboTwin the final student reaches 62.8% success compared to 63.6% for the teacher while using one fifth the accelerator time.
- Harness evolution raises teacher success from 33.4% to 63.6%
- Final 4B student outperforms open 4B and 27B agents
- Distilled students score close to teachers on training benchmarks
- Harness gains transfer to other frozen planners
- Student stays close to 552B planner on unseen benchmarks
From the original article
1 min read at huggingface.co
Robot agents built on vision-language models depend on both the planner and its harness: the tools, feedback, memory, and recovery logic around it. Can improvements to a large teacher's harness be carried into a compact agent with lower inference compute cost? We present TinyAstral, a two-stage approach.
By Kehong Gong, Xiaoxu Zheng, Guanli Hou · 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.