TinyAstral

Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents

With its weights frozen, a ~180B teacher improves through changes to its harness alone. Then a 4B student learns from its trajectories and takes its place in the same harness.

33.4→63.6
Teacher, weights frozen (%)
62.8vs63.6
4B student vs. ~180B teacher
6 benchmarks
3 training + 3 zero-shot
≈1/45
of the teacher's parameters
Huawei Ascend Atlas A2
Runs the real-robot planner
How it works

Evolve the harness. Distill the teacher. Deploy the student.

1 Evolve
Teacher ~180B · frozen harness H₁ Paired A/B Accept Readtrajectories Propose edit
33.4 → 63.6 teacher, weights unchanged
2 Distill
Teacher in final harness H* traces with reasoning + failure recovery 4B 0.8B–4B students collecttrain
62.8 4B student vs 63.6 teacher
3 Deploy
RoboTwin Guava LIBERO-PRO SimplerEnv VLABench EB-Manip. Real same harness H* 4B
≈1/5 of the teacher's accelerator time
Stage 1 · Harness RSI. LLM agents edit the frozen teacher's harness; a paired A/B test with a pre-registered gate decides what stays.
Stage 2 · Distillation. Freeze H*, then train 0.8B–4B students on teacher trajectories, reasoning and recovery included.
Deploy. The student replaces the teacher in the unchanged harness: 3 training + 3 zero-shot benchmarks, plus a real arm.
Paper figures Figure 1 (teaser) and Figure 2 (overview)
Figure 1: (a) harness RSI loop around a frozen ~180B teacher; (b) success on RoboTwin-50 across harness versions H1 to H*; (c) final students on six benchmarks.
Figure 1. (a) With teacher weights frozen, RSI improves the harness; we then freeze the final harness, collect teacher trajectories, and distill small students. (b) Harness evolution improves the teacher and other frozen planners; students trained separately at four harness versions each approach their corresponding teacher. (c) Final students on training and unseen benchmarks.
Figure 2: Stage 1 harness RSI around a frozen teacher; Stage 2 freeze H*, collect teacher traces, distill tiny students; deploy in the same harness on six benchmarks and a real robot.
Figure 2. Overview. Stage 1 evolves the harness around a frozen ~180B teacher: frontier LLM agents propose edits, and an edit is kept only if a paired A/B test passes a pre-registered gate. Stage 2 freezes the final harness H*, collects teacher traces (including recovery) plus offline real-robot supervision, and trains 0.8B–4B students that replace the teacher in the same harness, evaluated on three training domains, three unseen benchmarks and three real-robot tasks.
Real robot

TinyAstral-4B on a real arm, closed loop

Served on Huawei Ascend Atlas A2

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.

RSI evolution

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.

Strict success on RoboTwin-50 (%)
Teacher +30.2 pp and DeepSeek-V4.1-Flash +30.6 pp from H1 to H*; within H*, distillation adds 21.6 pp over the untrained 4B model.
Capability star map

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.

Six benchmarks

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).

Results

Main results

Success (%) on six benchmarks. 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 Final-state success only, without requiring a successful finish declaration. f 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. † >5% unknown outcomes, counted as failures. ✓/✗: condition satisfied / not satisfied; –: unavailable.

Student size vs. success
TinyAstral-0.8B / 2B / 4B in H*; dashed line: the ~180B teacher in the same harness.
Deployment cost
Planner accelerator·s per episode on 20 shared SimplerEnv episodes (median wall-clock time × planner accelerators, including simulation and perception; all planners on the same accelerator type). TinyAstral-4B uses about one fifth of its teacher's accelerator time.
Cite

BibTeX