TinyAstral improves a robot agent's harness, meaning the tools, feedback, memory and recovery logic around its planner, then distills the result into a small 4B model that keeps most of the gains.
Harness evolution: Frontier LLM agents read the teacher's trajectories and propose harness edits, and an edit is kept only if it passes a paired statistical test. On RoboTwin, the teacher's success rises from 33.4% to 63.6%.
Student model: The final 4B student reaches 62.8% on RoboTwin, and the paper says it stays close to its teacher on its three training benchmarks.
Compute: On the measured deployment episodes, the student uses about one fifth of its teacher's accelerator time.
Release: The submitter's comment says code, weights and data are released, and that the student runs closed loop on a real arm.
Sign in to comment.
TL;DR of the paper, generated automatically.
TinyAstral improves a robot agent's harness, meaning the tools, feedback, memory and recovery logic around its planner, then distills the result into a small 4B model that keeps most of the gains.