LLMs Edit Robot Policy Weights Directly: 2,048 Parameter Tweak Lifts Success from 68% to 88%
zhaoran_wang · x · 2026-11-07
Playground research released a first look at "weight-space programming" for robotics: frontier LLM agents (Claude Opus 5.5 and GPT-6 Astra) edited robot policies by writing weights directly, with no gradient descent, no new demonstrations, and no simulator in the loop.
Key findings:
- On π0.5, an edit touching just 2,048 weights (0.00006% of parameters) raised success on a perturbed LIBERO task from 68% to 88% — the edited policy learned to carry a resized mug lower and complete the task where the original policy wedged
- On small policies, agents removed learned failure habits and taught new skills by rewriting a handful of neurons
- Astra also built neural networks from scratch for cart-pole swing-up, cube pushing, and peg-in-hole insertion
The team suggests frontier LLMs may be able to write robot-control policies without collecting experience, while cautioning these are early results with small samples and limited simulated tasks.
More from Embodied
- InstantHMR open-sourced: real-time 3D human mesh recovery in 5.5ms per person on phone NPU — PurzBeats · 2026-11-07
- Replacing hand-tuned RL rewards with a decision model to teach humanoids to walk — tw_killian · 2026-11-07
- Muse starts rolling out to Meta Glasses, targeting hands-free AI agent control — armand_ruiz · 2026-11-07
- "2050 arrived early": a wave of humanoid robots on display — CyberRobooo · 2026-11-07
- "A year or two from now you'll just tell Astra 9 to make your humanoid do it," says SF builder — carlosdponx · 2026-11-07
- Open-source humanoid Asimov 1 ships as $20,000 DIY kit with ~20-hour build — freelerobot · 2026-11-07