TinyAstral in 85 seconds
Evolve the harness. Distill the teacher. Deploy the student.
Paper figures Figure 1 (teaser) and Figure 2 (overview)
Robot agents built on vision–language models depend on their harness: the tools, feedback, memory, and recovery logic around the planner. TinyAstral first freezes an open-weight ~180B teacher and improves only its harness through recursive self-improvement: frontier LLM agents propose edits, and each edit is kept only if it passes a paired statistical test. On RoboTwin this raises the teacher's success from 33.4% to 63.6%.
We then distill the teacher's trajectories, including reasoning and failure recovery, into a 4B student that replaces it in the same harness. The student reaches 62.8%, stays close to its teacher on three training benchmarks and to a 552B planner on three unseen ones, uses about one fifth of the teacher's accelerator time, and runs closed loop on a real SO-ARM101 arm served on Huawei Ascend Atlas A2.
TinyAstral-4B on a real arm, closed loop
The same harness drives a real SO-ARM101 arm: the 4B planner, served on Huawei Ascend Atlas A2, sees the cameras, calls one tool at a time and reads each result. Below: sorting two balls into matching boxes, and stacking three tape rolls.
Harness evolution lifts every planner we tested
Accepted harness versions measured on the same 500 RoboTwin-50 episodes. The first three planners keep their weights fixed; the student is distilled separately at each version.
Key evolution steps
Each step: an episode that fails under the parent harness, the accepted change, and the same kind of episode under the child harness.
One student, six benchmarks
Three benchmarks were used for RSI and training; three were never used for either. Click a card to watch TinyAstral-4B with the final harness H* (original evaluation runs, no re-runs).
Main results
Success (%) on six simulation benchmarks for embodied robot manipulation: RoboTwin-50, Guava and LIBERO-PRO (training), and SimplerEnv, VLABench and EB-Manipulation (zero-shot). Training benchmarks use strict success (autonomous done and terminal success); zero-shot benchmarks use their official measures. Small numbers under a cell are successes / episodes.
a Original LIBERO suites, single seed. b Ground-truth coordinates, boxes, and in-context examples. c,d Stop on success. e Position and task perturbations only (360 episodes), simulator point clouds; the student is trained on rollouts of the same tasks and tested on unseen seeds. f Official code and settings (per-step success flag, 90-minute budget) on our episodes. † >5% unknown outcomes, counted as failures. ✓/✗: condition satisfied / not satisfied; –: unavailable.