wheelleggedgenesis
June 27, 2025 · View on GitHub
Reinforcement learning of wheel-legged robots based on Genesis

sim2sim in mujoco

System Requirements
Ubuntu 20.04/22.04/24.04
python >= 3.10
Hardware requirements
NVIDIA/AMD GPU or CPU
must(必须)
It is not possible to install the 0.2.1 version of genesis with pip, with the local installation of the main branch because there is an API update
不能使用pip安装genesis的0.2.1版本,使用main分支的本地安装,因为有API更新
Test & Development Platform
- i5 12400f + Geforce RTX4070
- i7 12700kf + Radeon Rx7900xt
Before running
1. Clone repo
run:
It is recommended to use the latest release version here, the main branch of clone is not necessarily stable
这里建议使用最新的release版,clone的main分支不一定稳定
git clone https://github.com/Albusgive/wheel_legged_genesis.git
cd wheel_legged_genesis
2. install deps
use pdm install
Install pdm, https://pdm-project.org/en/latest/#installation, then run
pdm install
or manual install
Install Genesis:
https://github.com/Genesis-Embodied-AI/Genesis
install tensorboard:
pip install tensorboard
pip install pygame
pip install opencv-python
install rsl-rl:
cd rsl_rl && pip install -e .
Use
use pdm
test:
pdm run locomotion/wheel_legged_eval.py
train:
pdm run locomotion/wheel_legged_train.py
or manual
test:
python locomotion/wheel_legged_eval.py
train:
python locomotion/wheel_legged_train.py
gamepad & keyboard
gamepad
| key | function |
|---|---|
| LY | lin_vel |
| RX | ang_vel |
| LT | leg_length_up |
| RT | leg_length_down |
| LB | left_leg_length_ctrl_flag |
| RB | right_leg_length_ctrl_flag |
| X | Reset |
| XX | 太空步 |
| YY | 铁山靠 |
| B | 铁山靠回正 |
keyboard
| key | function |
|---|---|
| W | 前进 (Forward) |
| S | 后退 (Backward) |
| A | 左移 (Move Left) |
| D | 右移 (Move Right) |
| Q | 左转 (TURN Left) |
| E | 右转 (TURN Right) |
| Space | 站立 (up) |
| C/Left_Ctrl | 下蹲 (down) |
| 1 | left_leg_length_ctrl_flag(仅左腿) |
| 2 | right_leg_length_ctrl_flag(仅右腿) |
| Shift | 静步(Quiet Walking) |
| R | 重置环境(Reset) |
| UP&DOWN | 太空步 |
| LEFT&RIGHT | 铁山靠 |
| BACKSPACE | 铁山靠回正 |
Terrain
You can use the terrain as agent_eval_gym/agent_train_gym/agent_eval_gym/circular
| terrain | description |
|---|---|
| agent_train_gym | Rough roads and continuous slopes |
| agent_eval_gym | The agent_train_gym of the lite version |
| circular | Circular continuous slopes |
About custom terrains
Please refer to the code in assets/terrain
Suggestion
When using NVIDIA GPUs, it is recommended that the genesis backend choose gpu or cuda
When using AMD GPUs, it is recommended to select vulkan for the genesis backend
Demo of the effect
【开源】genesis双轮足机器人强化学习(基础版)
【开源】genesis轮足框架更新,增强地形适应能力
【开源】genesis强化学习仿真迁移测试
【开源】强化学习双轮足完美战胜mujoco
Changelog
| version | description |
|---|---|
| v0.0.1 | Basic functions and efficient training |
| v0.0.2 | add domain rand and curriculum |
| v0.0.3 | Faster and more stable training than v0.0.2 |
| v0.0.5 | Terrain training, custom terrain,add keyboard and better curriculum |
| v0.0.7 | sim2sim is a perfect victory over mujoco |
TODO:
- Rough roads:uphill and downhill slopes
- Curriculum
- Custom terrain
- Interference from external forces
- Sim2Sim:mujoco
- left_hip and right_hip
- swing legs & tie shan kao
- high speed control
技术交流:
微信超200人了,加QQ群吧
