TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents
TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents
This paper presents TinyAstral, a two-stage approach that first improves a large teacher model's harness through recursive self-improvement (RSI) and then distills the enhanced capabilities into a compact student agent. On RoboTwin, this method raises the teacher's success rate from 33.4% to 63.6%, with the final 4B student reaching 62.8%. The harness gains transfer to other frozen planners without retraining, and the student outperforms open 4B agents and a 27B harness-RSI model while using about one-fifth of the teacher's accelerator time.
The study introduces TinyAstral, a two-stage method that improves a large teacher model's harness via recursive self-improvement and distills the enhanced capabilities into a compact 4B student agent, achieving success rates near the teacher on RoboTwin while using approximately one-fifth of the accelerator time.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ough 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 succ
Open source S3
pen 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.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org