ONNX Policy Zoo

Open-sourced ONNX neural network policy files for humanoid robots

Overview

This repository provides pre-trained reinforcement learning controllers exported in ONNX format for direct deployment on humanoid robots. Policies cover locomotion, whole-body motion tracking, and combat/sports motions.

Supported Robots

Robot	Manufacturer	DOF
G1	Unitree Robotics	29
X2	Agibot	29

Repository Structure

onnx_policy/
├── unitree/
│   └── g1/
│       ├── unitree_locomotion/   # Locomotion policy with metadata
│       ├── whole_body_tracking/  # Motion tracking (dancing, idle, kicks, ...)
│       ├── boxing/               # Boxing and MMA strike motions
│       ├── legged_lab/           # AMP-based locomotion
│       └── holosoma_locomotion/  # Holosoma fastsac locomotion variants
└── agibot/
    └── x2/
        ├── whole_body_tracking/  # Motion tracking (dancing, idle, look around, ...)
        ├── boxing/               # Boxing and MMA strike motions
        └── holosoma_locomotion/  # Holosoma locomotion

Getting Started

Prerequisites

• Git LFS — all .onnx files are stored via Git LFS.
• ONNX Runtime — to run inference with the policies.

Clone

git lfs install
git clone https://github.com/IO-AI-TECH/onnx_policy.git

Running Inference

Each policy is a standard ONNX model. Load it with ONNX Runtime:

import onnxruntime as ort
import numpy as np

session = ort.InferenceSession("unitree/g1/unitree_locomotion/g1_policy.onnx")
input_name = session.get_inputs()[0].name

obs = np.zeros((1, 47), dtype=np.float32)  # num_observations = 47
action = session.run(None, {input_name: obs})[0]

Refer to the accompanying *_metadata.json for the correct observation dimension, joint order, PD gains, and scaling factors for each policy.

Policy Metadata

For policies that include a metadata file, the fields are:

Field	Description
num_observations	Observation vector dimension
num_actions	Number of controlled joints
action_scale	Scalar multiplier applied to policy output
joint_names	Ordered joint names matching action indices
default_joint_pos	Nominal joint positions (radians)
kp / kd	PD stiffness and damping gains per joint
observation_names	Ordered observation components
*_scale	Scaling factors for each observation component
is_recurrent	Whether the policy uses an LSTM
rnn_type / rnn_hidden_size / rnn_num_layers	RNN architecture (if recurrent)

License

This project is licensed under the MIT License.