unitree-go2-velocity-flat

diasAiMaster/unitree-go2-velocity-flat

unitree-go2-velocity-flat is a reinforcement learning model from diasAiMaster. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. It is set up for rsl-rl. The card lists the license as bsd-3-clause.

RL locomotion policy for the Unitree Go2 quadruped robot, trained on flat terrain using PPO.

Downloads · 30 days

Access

Public

Updated May 10, 2026

Repo size

29.3 MB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.pt27.4 MB · 84%

At a glance

Task: Reinforcement Learning
Library: rsl-rl
License: bsd-3-clause
Access: Public
Created: Feb 6, 2026
Updated: May 10, 2026
SHA: 9a723b7a

Model card

From the Hugging Face model README

Unitree Go2 — Velocity Flat (PPO)

RL locomotion policy for the Unitree Go2 quadruped robot, trained on flat terrain using PPO.

Demo

Training

Framework: unitree_rl_mjlab (MuJoCo Warp)
Task: Mjlab-Velocity-Flat-Unitree-Go2
Algorithm: PPO (RSL-RL)
Hardware: 10× NVIDIA RTX A4000, 56 CPU cores
Environments: 8192 parallel
Training time: ~18 minutes (506 iterations)

Results

Metric	Value
Mean reward	52.9
Mean episode length	1000 (max, no falls)
Steps/sec	628K-738K

Files

File	Description
`policy.onnx` + `policy.onnx.data`	ONNX model for deployment (go2_ctrl)
`model_500.pt`	Final PyTorch checkpoint (best for fine-tuning)
`model_0.pt` ... `model_400.pt`	Intermediate checkpoints every 100 steps
`params/deploy.yaml`	Deploy configuration (obs order, action scale, joint mapping)
`params/env.yaml`	Environment configuration
`params/agent.yaml`	Agent/PPO configuration
`events.out.tfevents.*`	TensorBoard training logs

Usage

Deploy in MuJoCo simulator

# Copy ONNX model + deploy config
cp policy.onnx policy.onnx.data \
  unitree_rl_mjlab/deploy/robots/go2/config/policy/velocity/v0/exported/
cp params/deploy.yaml \
  unitree_rl_mjlab/deploy/robots/go2/config/policy/velocity/v0/params/

# Build controller
cd unitree_rl_mjlab/deploy/robots/go2
mkdir -p build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release && make -j$(nproc)

# Run simulator + controller
cd unitree_mujoco/simulate/build && ./unitree_mujoco
cd unitree_rl_mjlab/deploy/robots/go2/build && ./go2_ctrl --network=lo

Important: This model was trained without gait_phase and with action scale 0.5. The default deploy.yaml in unitree_rl_mjlab may differ — use params/deploy.yaml from this repo.

Fine-tune on rough terrain

# Place model_500.pt in logs/rsl_rl/go2_velocity/<run_name>/
python scripts/train.py Mjlab-Velocity-Rough-Unitree-Go2 \
  --agent.resume=True \
  --agent.load-run="<run_name>" \
  --agent.load-checkpoint="model_500.pt" \
  --agent.algorithm.learning-rate=1e-4

Known Issues

The upstream unitree_rl_mjlab has bugs that crash multi-GPU training on rough terrain — see Issue #9 and PR #8.

Task: Reinforcement Learning
Library: rsl-rl
License: bsd-3-clause
Created: Feb 6, 2026
Updated: May 10, 2026