Papers.Data.Code
Paper

PhysEvo: Astra Can Act, Let It

Wenqing Tian, Zeyu Zhang, Zhaocheng Liu, Fengwei Liu, Qiang Liu, Liang Wang

A frozen-model RSI loop that edits robot control, observation, and skills from trajectory evidence, then validates and retains fixes for later tasks.

arXiv ↗Oct 6, 2026

  • robotics
  • llm
  • robot manipulation
  • agentic systems
  • self-improvement
Pipeline diagram showing Astra's task/meta agents, harness evolution, and redeployment loop
Pipeline diagram showing Astra's task/meta agents, harness evolution, and redeployment loop

What it is

A frozen robot foundation model improves itself by editing its action/observation harness and diagnostic tools from execution traces, instead of updating weights or training a policy. Instead of retraining a robot policy, execution failures are turned into persistent harness edits: the same frozen model acts, diagnoses trajectories, revises tools/skills, tests fixes, and reuses improved diagnostics.

Key results

  • 62.0% success on 42 RoboDojo tasks vs 47.17% for RoboDawn
  • 55.0% success on 8 hard manipulation tasks vs 1.25% direct Astra
  • 84.0% success and 90.6/100 on 5 real-world tasks over 25 trials

Reliable robot behavior can be improved substantially by evolving the harness around a frozen model, not the model weights.

One useful ML email a week.

More from Oct 12 – Oct 18

The whole week →

Picked Oct 16

Picked Oct 15

  • Repo
    ziplab/ATAS

    Adds explicit scene and event memory with validated world updates, targeting spatial revisit and event recapture beyond standard multi-shot generation.

    GitHubOct 12, 2026
    →
  • Dataset
    HelenPeng/ExeWorldBench

    Targets executable Blender world construction, scoring functional correctness instead of visual similarity alone.

    Hugging FaceOct 9, 2026
    →

Picked Oct 14

  • Dataset
    ScaleAI/HSS

    Makes intuitive visual reasoning measurable with criterion-level grading; humans hit 93.1% vs 53.6% for the best tested model.

    Hugging FaceOct 1, 2026
    →

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.