Compact robot agents match large planners with self improved harnesses

TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents

RoboticsComputer Vision and Pattern Recognition

Summary

Robot agents that use vision and language models rely heavily on a supportive framework called a harness, which includes tools and strategies for planning and correcting actions. The authors show how improvements made to a large, powerful robot planner’s harness can be transferred to a much smaller, faster robot. They do this by letting language models suggest improvements to the harness, testing those improvements, then training a smaller robot to use the improved harness and mimic the large planner’s behavior. The smaller robot achieves nearly the same success as the bigger one but runs much faster.

What this means in practice

  • For industrial robot integrators: Deploy compact robot agents that perform complex manipulation tasks nearly as well as large planners but with lower compute cost using self-improved harnesses.
  • For mobile robot developers: Build small robot agents that adapt and recover from failures effectively by inheriting harness improvements from larger teacher models without retraining.

Authors

Kehong Gong, Xiaoxu Zheng, Guanli Hou, Seunghun Lee, Mengqi Guo, Runjia Tan, Zhengyu Wen, Qi Wang, Xiaoyu He, Mingyuan Zhang, Michael Bi Mi

Abstract

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. We first freeze an open-weight teacher and improve its harness through recursive self-improvement (RSI): frontier LLM agents read the teacher's trajectories and propose harness edits, and each edit is accepted only if it passes a paired statistical test. We then freeze the final harness, collect teacher trajectories with reasoning and failure recovery, and distill a small student that replaces the teacher in the same harness. On RoboTwin, harness evolution raises the teacher's success from 33.4% to 63.6%; the final 4B student reaches 62.8%. Across four harness versions, distilled 4B students score close to their corresponding teachers, and later versions yield stronger students, showing that harness evolution gains carry into compact models. The harness gains also transfer to other frozen planners without retraining. On identical episodes, the final 4B student outperforms open 4B agents, the open 27B harness-RSI model SenseNova-RoboRSI, and our own teacher model placed in three competitors' harnesses. It stays close to its teacher on all three training benchmarks, and on three benchmarks never used for RSI or training it scores close to a 552B planner running the same harness. On the measured deployment episodes, it uses about one fifth of its teacher's accelerator time.