← 开源
mit-han-lab

pearNEW

Training Reasoning Models in Simulation to Natively Use Real Robots

Python
在 GitHub 打开
增长势头
+5124 小时新增 Star+7.7%
712
Star
45
Fork
3
贡献者
创建于 2026-11-13 · 更新于 2026-11-19 · 今日第 370 名
主要开发者
SakitsJerryGJXzhuoyang20
README

a Post-training Engine for Agentic Robotics

| Blog | Paper | Model | Documentation |


About

PEAR is a post-training engine for agentic robotics: it trains reasoning models to natively use robots, with reinforcement learning in simulation.

PEAR is the training infrastructure behind Training Reasoning Models in Simulation to Natively Use Real Robots, where an open 27B reasoning model learns to use real robots entirely in simulation, with zero demonstrations.

PEAR trains with:

  • Fully asynchronous RL on veRL
  • GRPO, dynamic sampling, and Score Centering
  • FSDP and sequence parallelism for long-context training
  • Memory-efficient training with gradient checkpointing and activation offloading
  • Prefix caching across rounds, with LMCache offloading the KV cache to CPU memory
  • Optimized kernels for H100, H200 and GB300
  • Simulators on the same Ray cluster: MuJoCo on CPUs, Isaac Sim on GPUs

PEAR evaluates with:

  • Robotic benchmarks: LIBERO, LIBERO-Plus, LIBERO-Pro, RoboDojo and RoboLab-120
  • General benchmarks: nine reasoning, coding and vision benchmarks through EvalScope

PEAR deploys with:

  • A model server: vLLM server on one RTX 5090 or GB300, with decode megakernels
  • A robot station: the agent loop that drives a YAM or an SO-101 robot

News

  • [2026/11] 🔥 Check out our latest DFlash2 checkpoint for Qwen3.8-27B-PEAR! With it, the NVFP4 model decodes at 921 tokens/s on one GB300 and 262 tokens/s on one RTX 5090!

Getting Started

Install PEAR and everything a LIBERO run needs, on a GPU node:

git clone https://github.com/mit-han-lab/pear.git && cd pear
export WORK_DIR=<a directory every node sees>   # the model, the environments and the runs go here
bash scripts/install.sh libero_run              # installs the model, LIBERO and the training environment

Then start a Ray cluster and launch your first run:

Supported Models

Model Recipes
Qwen3.8-27B train, deploy

Supported Environments

Environment Robot Simulator Used for
LIBERO Franka MuJoCo training, evaluation
LIBERO-Plus Franka MuJoCo evaluation
LIBERO-Pro Franka MuJoCo evaluation
RoboDojo dual-arm ARX Isaac Sim training, evaluation
RoboLab-120 Franka with a Robotiq gripper Isaac Sim evaluation

Supported Embodiments for Deployment

Robot Arms Used for
YAM two, with grippers deployment
SO-101 one, with a gripper deployment

Supported Hardware

Platform Hardware
NVIDIA Blackwell GB300, RTX 5090 (the model server)
NVIDIA Hopper H100, H200

Acknowledgment

PEAR is built upon veRL and LMCache, and uses vLLM, and Ray.

排名相邻

AI Rank iPhone 版

随时追踪 AI 模型榜单

在 App Store 下载