
MIT unveils PEAR for agentic robotics
MIT Han Lab, alongside UC Berkeley, UCSD, and NVIDIA, released PEAR, an open-source post-training engine that uses reinforcement learning in simulation to train reasoning models for robotics without human demonstrations. An open-weight 27B model trained with PEAR jumps from 1.6% to 74.4% on LIBERO while retaining full vision and coding capabilities.
Imitation learning created brittle robots that mimic demonstrations without understanding intent, but PEAR proves reasoning models trained purely via sim RL can actually generalize to new instructions. Treating robot manipulation as an end-effector tool-use problem mirrors how computer-use agents evolved, bypassing the multi-million-dollar teleoperation data moat entirely.
- –Zero-shot sim-to-real transfer: By optimizing on pure task-success signals across physics simulators like MuJoCo and Isaac Sim, models transfer to real hardware with a 60% success rate without physical demonstrations.
- –Generalization over imitation: Under LIBERO-Pro perturbations where tasks or layouts shift, demonstration-based models like π0.5 plummet from 96.9% to 0.75%, whereas PEAR's reasoning agent retains a 57.4% success rate.
- –Preserved general intelligence: Unlike behavioral cloning runs that cause catastrophic forgetting, models post-trained with PEAR preserve baseline performance across nine reasoning, vision, and coding benchmarks including AIME and LiveCode.
- –Efficient local deployment: Built on asynchronous RL (veRL) with GRPO, KV cache CPU offloading via LMCache, and vLLM decode kernels, allowing inference on local hardware like single RTX 5090 or GB300 accelerators.
DISCOVERED
39m ago
2026-11-16
PUBLISHED
1h ago
2026-11-16
RELEVANCE
AUTHOR
ajcyril