Embodied Models

Choose a model or policy class to find its installation, checkpoint, and RL training workflow. For examples organized by benchmark, see RL with Embodied Simulators.

Hardware Support

NVIDIA is the default backend in the model recipes. AMD ROCm, Huawei Ascend CANN, and Moore Threads MUSA also support the three model families below on LIBERO and ManiSkill. Follow a model link for installation and launch instructions. On non-CUDA backends, ManiSkill runs PhysX on CPU and selects the renderer independently.

Model	Hardware	Environment and Scope
OpenVLA-OFT	AMD ROCm · Huawei Ascend CANN · Moore Threads MUSA	LIBERO · ManiSkill
GR00T N1.5	AMD ROCm · Huawei Ascend CANN · Moore Threads MUSA	LIBERO · ManiSkill with a maniskill_widowx checkpoint
π₀ / π₀.₅ (OpenPI)	AMD ROCm · Huawei Ascend CANN · Moore Threads MUSA	LIBERO · ManiSkill · RoboTwin on ROCm

Model Recipes

Open a model page for the full workflow and its supported environments.

OpenVLA-OFT
LIBERO training on NVIDIA CUDA, AMD ROCm, Huawei Ascend CANN, and Moore Threads MUSA

RL on MLP Policy
Train a lightweight MLP policy with PPO, SAC, or GRPO across simulation environments

RL on π₀-FAST
LeRobot π₀-FAST + LIBERO-Long + token-level GRPO

RL on π₀ and π₀.₅ Models
Significant improvement in RL training on π₀ and π₀.₅

RL on GR00T Models
Support GR00T-N1.5, N1.6 and N1.7 RL fine-tuning.

RL on Dexbotic Model
Dexbotic (π₀.₅-based) + LIBERO + PPO training

RL on StarVLA Models
StarVLA + LIBERO + GRPO embodied RL training

MolmoAct2 Evaluation
Evaluate the official MolmoAct2-LIBERO checkpoint on LIBERO

RL with LingbotVLA Models
Support LingbotVLA 1.0 (GRPO) and 2.0 (PPO) + RoboTwin

RL on Evo-1 Models
Train the Evo-1 vision-language-action model with embodied RL

RL on ABot-M0 Model
ABot-M0 native integration with LIBERO-plus PPO training

RL on Psi0 Model
Psi0 + SIMPLE + PPO humanoid RL training