---
title: tron1-rl-locomotion-policies
canonical_url: "https://www.modelscope.cn/models/LimXDynamics/tron1-rl-locomotion-policies"
md_url: "https://www.modelscope.cn/models/LimXDynamics/tron1-rl-locomotion-policies.md"
repository: LimXDynamics/tron1-rl-locomotion-policies
last_updated: 2026-11-16
license: apache-2.0
pipeline_tag: robotics
tasks:
  - robotics
library_name:
  - onnx
  - pytorch
frameworks:
  - pytorch
language:
  - en
  - zh
downloads: 15
stars: 0
tags:
  - robotics
  - reinforcement-learning
  - legged-robots
  - biped
  - locomotion
  - loco-manipulation
  - onnx
  - isaac-gym
  - isaac-lab
  - mujoco
  - limx-dynamics
  - tron1
---

# tron1-rl-locomotion-policies

> tron1-rl-locomotion-policies - LimXDynamics 在 ModelScope 开源的模型。Try it in your browser / 在浏览器里试试: TRON 1 Teleop runs the point-foot, sole-foot and wheel-foot policies from this repo (PFTRON1A, SFTRON1A, WFTRON1A) in MuJoCo, right in the browser. Drive it with…

LimXDynamics/tron1-rl-locomotion-policies 是 ModelScope 魔搭社区上的robotics模型，采用 apache-2.0 许可。

- **Repository**: LimXDynamics/tron1-rl-locomotion-policies
- **License**: apache-2.0
- **Tasks**: robotics
- **Tags**: robotics, reinforcement-learning, legged-robots, biped, locomotion, loco-manipulation, onnx, isaac-gym, isaac-lab, mujoco, limx-dynamics, tron1
- **Downloads**: 15
- **Stars**: 0
- **Last updated**: 2026-11-16

Source: https://www.modelscope.cn/models/LimXDynamics/tron1-rl-locomotion-policies

---

[![Open in Spaces](https://huggingface.co/datasets/huggingface/badges/resolve/main/open-in-hf-spaces-md.svg)](https://huggingface.co/spaces/limxdynamics/tron1-teleop)

**Try it in your browser / 在浏览器里试试**: [TRON 1 Teleop](https://huggingface.co/spaces/limxdynamics/tron1-teleop) runs the point-foot, sole-foot and wheel-foot policies from this repo (`PF_TRON1A`, `SF_TRON1A`, `WF_TRON1A`) in MuJoCo, right in the browser. Drive it with W A S D, no install. 打开网页就能用键盘遥控 TRON 1，点足、双足、轮足三种形态随时切换，跑的就是本仓库的策略，无需安装。

## 中文说明

TRON 1 全部三种形态（点足 PF、双足 SF、轮足 WF）的强化学习行走策略，ONNX 格式。每个型号都有 Isaac Gym 和 Isaac Lab 训练的两版策略，可在 CPU 上用 ONNX Runtime 运行，支持 MuJoCo 仿真和真机部署，配套部署代码为 [tron1-rl-deploy-python](https://github.com/limxdynamics/tron1-rl-deploy-python)。

`arm/` 目录里另有 TRON 1 加装 6 自由度机械臂的全身策略（双足 `SF_TRON1A`、轮足 `WF_TRON1A`）：一个网络同时输出 8 个腿部关节和 6 个手臂关节的目标，边走边操作，配套部署代码为 [tron1-rl-deploy-arm](https://github.com/limxdynamics/tron1-rl-deploy-arm)（ROS Noetic + Gazebo）。

文件原样同步自 GitHub，以 GitHub 仓库为准。

### 快速上手（仿真）

1. 克隆部署代码 [tron1-rl-deploy-python](https://github.com/limxdynamics/tron1-rl-deploy-python) 和仿真器 [tron1-mujoco-sim](https://github.com/limxdynamics/tron1-mujoco-sim)（部署代码已包含这些策略）
2. 安装 limxsdk 以及 numpy、pyyaml、scipy、onnxruntime、mujoco
3. 设置环境变量 `ROBOT_TYPE`（如 `PF_TRON1A`）和 `RL_TYPE`（`isaacgym` 或 `isaaclab`）
4. 分别启动 `simulator.py` 和 `main.py`，在 MuJoCo 窗口按 Backspace 重置，机器人会站起并原地踏步

只下载某一个型号，例如点足 TRON 1A：

```bash
hf download limxdynamics/tron1-rl-locomotion-policies --include "PF_TRON1A/*" --local-dir ./tron1-policies
```

国内用户也可以从魔搭下载：[LimXDynamics/tron1-rl-locomotion-policies](https://www.modelscope.cn/models/LimXDynamics/tron1-rl-locomotion-policies)。

真机部署前请先在仿真中测试，确保有经过培训的操作人员在场，急停随手可及。

采购咨询与合作：**opensource@limxdynamics.com**

---

# TRON 1 RL Locomotion Policies

![TRON 1 point-foot, sole-foot and wheel-foot](https://raw.githubusercontent.com/limxdynamics/tron1-rl-deploy-python/main/doc/banner.png)

Reinforcement-learning locomotion policies in ONNX for the **LimX TRON 1** biped in all three forms: point-foot (`PF_*`), sole-foot (`SF_*`) and wheel-foot (`WF_*`). Every model has a policy trained in Isaac Gym and one trained in Isaac Lab. They run on CPU with ONNX Runtime, in MuJoCo or on the real robot, through [tron1-rl-deploy-python](https://github.com/limxdynamics/tron1-rl-deploy-python).

These files are mirrored as-is from [tron1-rl-deploy-python](https://github.com/limxdynamics/tron1-rl-deploy-python/tree/main/controllers/model), and the arm policies under `arm/` from [tron1-rl-deploy-arm](https://github.com/limxdynamics/tron1-rl-deploy-arm/tree/main/src/robot_controllers/config/pointfoot). The GitHub repositories are the source of truth.

![TRON 1 walking in MuJoCo](https://raw.githubusercontent.com/limxdynamics/tron1-rl-deploy-python/main/doc/simulator.gif)

## What's included

| Family | `ROBOT_TYPE` folders | Joints / actions |
| --- | --- | --- |
| Point-foot | `PF_TRON1A`, `PF_TRON1B`, `PF_P441A`, `PF_P441B`, `PF_P441C`, `PF_P441C2` | 6 / 6 |
| Sole-foot | `SF_TRON1A`, `SF_TRON1B` | 8 / 8 |
| Wheel-foot | `WF_TRON1A`, `WF_TRON1B` | 8 / 8 |

Each folder holds `params.yaml` and a policy pair at `policy/<RL_TYPE>/policy.onnx` and `policy/<RL_TYPE>/encoder.onnx`, where `RL_TYPE` is `isaacgym` or `isaaclab`. The control loop runs at 500 Hz and the policy at 50 Hz.

### Arm-equipped TRON 1 (loco-manipulation)

| Model | Folder | Joints / actions |
| --- | --- | --- |
| Sole-foot + 6-DoF arm | `arm/SF_TRON1A` | 14 / 14 (8 legs + 6 arm) |
| Wheel-foot + 6-DoF arm | `arm/WF_TRON1A` | 14 / 14 (8 legs + 6 arm) |

One network walks the robot and moves the arm (joints J1–J6) at the same time, so the body stays balanced while the arm reaches. Each folder holds `params.yaml`, `policy/encoder.onnx` (the last 5 observations, 405 → 64) and `policy/policy.onnx` (81 + 64 → 14 joint targets). Deploy them with [tron1-rl-deploy-arm](https://github.com/limxdynamics/tron1-rl-deploy-arm) (ROS Noetic, Gazebo), which already contains these files and the arm-equipped robot models.

```bash
hf download limxdynamics/tron1-rl-locomotion-policies --include "arm/*" --local-dir ./tron1-policies
```

## Quick start (simulation)

Linux x86_64 or aarch64, Python 3, CPU only. The deploy repository already contains these policies.

```bash
git clone --recurse-submodules https://github.com/limxdynamics/tron1-rl-deploy-python.git
git clone --recurse-submodules https://github.com/limxdynamics/tron1-mujoco-sim.git
pip install tron1-rl-deploy-python/limxsdk-lowlevel/python3/amd64/limxsdk-*.whl
pip install numpy pyyaml scipy onnxruntime mujoco

export ROBOT_TYPE=PF_TRON1A
export RL_TYPE=isaacgym
python tron1-mujoco-sim/simulator.py      # terminal 1
python tron1-rl-deploy-python/main.py     # terminal 2
```

Press Backspace in the MuJoCo window to reset the robot; it stands up and walks in place. Drive it with `./tron1-mujoco-sim/robot-joystick/robot-joystick`.

To download only one model from here, for example the point-foot TRON 1A:

```bash
hf download limxdynamics/tron1-rl-locomotion-policies --include "PF_TRON1A/*" --local-dir ./tron1-policies
```

## Running on the robot

Pass the robot's IP to `main.py` (`python main.py <robot-ip>`). The controller waits for calibration state 0, starts on **L1 + Y** and stops on **L1 + X** (zero stiffness, damping only). Test in simulation first, run with a trained operator present, and keep the emergency stop within reach.

## Related

- Robot models (URDF / MJCF): [tron1-robot-description](https://github.com/limxdynamics/tron1-robot-description)
- Simulator: [tron1-mujoco-sim](https://github.com/limxdynamics/tron1-mujoco-sim)
- Training: [tron1-rl-isaacgym](https://github.com/limxdynamics/tron1-rl-isaacgym) · [tron1-rl-isaaclab](https://github.com/limxdynamics/tron1-rl-isaaclab)
- ROS deployment: [tron1-rl-deploy-ros](https://github.com/limxdynamics/tron1-rl-deploy-ros) · [tron1-rl-deploy-ros2](https://github.com/limxdynamics/tron1-rl-deploy-ros2)
- Arm deployment (loco-manipulation): [tron1-rl-deploy-arm](https://github.com/limxdynamics/tron1-rl-deploy-arm)

## License

Apache 2.0, the same as [tron1-rl-deploy-python](https://github.com/limxdynamics/tron1-rl-deploy-python).

## Contact

- Business inquiries and collaboration: **opensource@limxdynamics.com**
- Questions: open a discussion in the Community tab or an issue on GitHub
- More from LimX Dynamics: [Hugging Face](https://huggingface.co/limxdynamics) · [GitHub](https://github.com/limxdynamics) · [Website](https://www.limxdynamics.com/en)
