AI/ robotics · reinforcement-learning · ai · open-source

Robots Learn From Failure With Z-1 RL Framework

A new post-training method lifts robot task success rates by 13 points using only public data and no extra demonstrations.

A reinforcement learning framework called Z-1 teaches robots to get better by failing, not just by copying humans.

Most robot AI today learns by watching recorded demonstrations — a method called behavior cloning. Z-1, built on top of an existing model called π0.5, takes a different route: after an initial supervised training pass on publicly available RoboCasa kitchen task demos, it applies a technique called Group Relative Policy Optimization (GRPO) to let the model keep improving from its own attempts. The system uses four efficiency tricks — shared-prefix rollout construction, tree-structured trajectory branching, completion-aware reward calibration, and selective joint training — to keep that online learning stable without spiraling into chaos.

The result is an average 80.6% success rate across 24 standard RoboCasa tasks, a 13.2 percentage-point gain over the supervised-only baseline. The catch that isn't really a catch: Z-1 used only publicly released data, meaning labs without proprietary demonstration libraries can reproduce or build on it.

Reinforcement learning has long been the missing ingredient in physical robotics — it works brilliantly for games and simulations but tends to be brittle and slow when a real robot arm has to touch real objects. If GRPO-style post-training can close that gap reliably, the field may finally stop treating "more human demos" as the default answer to every performance plateau.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →