mit-han-lab/pear
Machine LearningPost-training engine that trains reasoning models to control real robots via reinforcement learning in simulation.
What PEAR Does
PEAR is a post-training engine for agentic robotics. It trains reasoning models to use robots natively through reinforcement learning in simulation, aiming to produce policies that transfer to real hardware without demonstration data. The reference result is an open 27B reasoning model that learns robot use entirely in simulation.
The training stack is built on veRL and offers fully asynchronous RL, GRPO, dynamic sampling, Score Centering, FSDP and sequence parallelism for long-context training, gradient checkpointing and activation offloading, prefix caching with LMCache, and optimized kernels for H100, H200 and GB300. Simulators run on the same Ray cluster: MuJoCo on CPUs and Isaac Sim on GPUs.
Evaluation covers robotic benchmarks (LIBERO, LIBERO-Plus, LIBERO-Pro, RoboDojo, RoboLab-120) and general reasoning, coding and vision benchmarks through EvalScope. Deployment combines a vLLM model server on an RTX 5090 or GB300 with a robot-station agent loop driving a YAM or SO-101 robot.
Quick Start
The documented path is a LIBERO training run on a GPU node. You need a directory visible to every node in the cluster, since the model, environments and run outputs all live there.
git clone https://github.com/mit-han-lab/pear.git && cd pear
export WORK_DIR=<a directory every node sees>
bash scripts/install.sh libero_run
The install script sets up the model, LIBERO and the training environment. Afterwards, start a Ray cluster and launch the run. The installation guide, quick start and example recipes cover the cluster setup and run commands that the README does not inline.
Supported Models and Environments
Qwen3.8-27B has both training and deployment recipes; the deployment recipe includes instructions for driving a robot example deploy README. Training and evaluation environments are LIBERO and RoboDojo (Franka/MuJoCo and dual-arm ARX/Isaac Sim respectively), while LIBERO-Plus, LIBERO-Pro and RoboLab-120 are evaluation-only. Deployment embodiments are YAM (dual arm with grippers) and SO-101 (single arm with gripper).
Hardware and Operational Constraints
Training targets NVIDIA Hopper (H100, H200) and Blackwell (GB300, RTX 5090) hardware; the model server runs on one RTX 5090 or GB300. The stack depends on veRL, LMCache, vLLM and Ray, and assumes a multi-node Ray cluster for the full setup. Isaac Sim environments need GPUs, while MuJoCo simulators run on CPUs. Real-robot deployment requires the listed YAM or SO-101 arms, and the repository is licensed under Apache-2.0.
HotRepo's take
PEAR entered HotRepo's new-project list on the Python list on 2026-11-14, gaining 2 stars in a roughly one-hour observation window to reach 103 stars and 10 forks. That is a short series with no full-language daily rank recorded, so it supports only the fact of early interest, not a trend. The substance behind it is a concrete training-and-deployment pipeline tied to a published paper and an open 27B model, which makes it most relevant to robotics researchers with Hopper/Blackwell clusters rather than casual users.
Written by AI from the project's README, updated . The project's own docs are the source of truth.
You might also like
- Open-RUA/uni-robot-agentRun robot-use agents in simulated manipulation environments via CLI, Python SDK, or local dashboard.
- pollen-robotics/microduck_rlTraining environments and sim2real recipes for teaching a small bipedal robot to walk with reinforcement learning.
- apirrone/Open_Duck_MiniOpen hardware plans, 3D files and reinforcement-learning code for building and training a small bipedal duck robot at home.
- ninetyhe-90/training-exploreInteractive lessons that let you play, step through, and verify a small Transformer's real forward-pass numbers.
- apple-aiml-research/ml-haloResearch code for generating synthetic data, training, and deploying perceptive humanoid motion policies for the Unitree G1.