AlphaNIST: binary-reward RL trains one policy to 98.2% success on all 20 NIST assembly tasks
chris_j_paxton · x · 2026-10-14
Inspired by AlphaGo/AlphaZero, AlphaNIST trains robots tabula rasa with only a binary success reward plus a simple control cost: a single policy solves all 20 NIST assembly tasks at 98.2% success in simulation and can complete 10 different tasks in a row. Notably, it's not scaling-pilled — decent success rates come from just 4k parallel environments and 25k PPO iterations (3.2B steps), with experiments spanning 1–256K environments.
Related event: NVIDIA's AlphaNIST Masters 20 Assembly Tasks with Binary Rewards(2 posts)→
More from Embodied
- FLUX 3 Action Tops RoboLab-120 at 42.9%, Bringing Video Pretraining to Robotics — DeepLearningAI · 2026-10-15
- DeployRobo launches as the generalist robot deployment company, backed by Y Combinator — ycombinator · 2026-10-14
- Robot foundation-model startups are all opening APIs — the model layer was never the moat — chris_j_paxton · 2026-10-14
- Metal Arm: a $2,499 6-axis CNC robot arm that ships in a week, built on LeRobot — chris_j_paxton · 2026-10-14
- NovaPlan (CoRL 2026): zero-shot long-horizon robot manipulation with no demonstrations — chris_j_paxton · 2026-10-14
- Ukrainian FPV Drone Sim Leaves Early Access, Players Generate Training Data — generativist · 2026-10-14