MIT team trains open 27B model with zero demos to score 18/30 on a real robot it never touched
songhan_mit · x · 2026-11-14
MIT Han Lab and collaborators present PEAR (Post-training Engine for Agentic Robotics): they post-train an open Qwen3.8-27B model purely through its own attempts in simulation, with zero human demonstrations.
- Results: on a real robot it had never touched, success jumps from 0 to 18/30 tasks; it can also copy a painting and design its own chemistry experiment
- Benchmarks: better generalization and instruction-following than π0.5; close to Fable's closed Astra max on LIBERO
- Deployment: runs locally at 220+ tok/s on an RTX 5090
- Notable: general capabilities like math and coding are preserved after post-training
Model weights, code, and full training infra are open-sourced.
More from Embodied
- A new wearable lets you vibe code from your wrist — ZeroStateReflex · 2026-11-14
- Melty Gym launches: train robots in any game at massive scale — julianweisser · 2026-11-14
- NVIDIA unveils GR00T 2, its most advanced open robotics model with cross-body skill transfer — nvidia · 2026-11-14
- Google's packed week: Gemini 4 Carbon, Decisions API, Robotics 2.5, AlphaGenome 2 — osanseviero · 2026-11-14
- China Took 59% of 600,000+ Industrial Robots Installed in 2025; Humanoid Shipments Still Tiny — ingliguori · 2026-11-14
- Y Combinator-backed Nines Robotics launches autonomous robots for AI data centers — ycombinator · 2026-11-14