Skip to content

About

No description, website, or topics provided.

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

OpenDoge Train(mjlab)

Hardware License

OpenDoge 是 OpenDogeRobotics 开源的四足机器人。本仓库 OpenDoge_train 负责它的强化学习训练、仿真验证与策略导出。

仓库有两条并行分支,训练栈不同、命令不通用:

  • mjlab(默认分支,本 README 描述的内容):基于 mjlab 1.6.0(MuJoCo-Warp)的当前主线,包含 OpenDoge 的速度/起身/倒立/跳跃/多技能任务、Viser 回放、独立工作台与浏览器试玩区。
  • isaacgym:原 Isaac Gym 训练栈(Preview 4 + legged_gym + HIMLoco 衍生的 PPO/AMP 实现),保留 MuJoCo Sim2Sim 验证与 ONNX 导出。用于复现历史结果,以及和 mjlab 版本做对照。
git checkout isaacgym   # 切到 Isaac Gym 版本
git checkout mjlab      # 切回当前主线

分支对照

分支 训练栈 主要内容 定位
mjlab(默认) Python 3.10–3.13 + mjlab 1.6.0 + MuJoCo-Warp + RSL-RL + PyTorch OpenDoge 与 Unitree 机型的任务注册、Viser 回放、独立工作台、多技能策略、ONNX 导出 当前开发主线
isaacgym Python 3.8 + Isaac Gym Preview 4 + legged_gym + HIMLoco rsl_rl OpenDoge/A1/Go1 训练配置、MuJoCo Sim2Sim 验证、ONNX 导出 历史版本,维护性修复

为什么用 mjlab

  • 不依赖 Isaac Gym:mjlab 是纯 Python + MuJoCo-Warp 实现,CPU 单环境即可跑通环境构建、reset 与一步动作,开发机不需要 GPU 和 Isaac Gym 安装包。
  • 资产与任务集中管理:机器人资产在 src/assets/,任务差异在 src/tasks/robots/,训练、回放、工作台共用同一套注册表。
  • 版本固定、适配层薄:mjlab==1.6.0 作为固定依赖,本仓库只维护机器人资产、任务差异、验证契约和部署适配,不复制框架代码。
  • 训练/验证/导出同一套 CLI:train、play、baseline-play、opendoge-eval、opendoge-export 都在仓库内。

机器人资产与任务

OpenDoge 是本仓库的主要机器人:

  • 速度跟踪:LainLab-OpenDoge-Flat、LainLab-OpenDoge-Rough
  • 单技能:LainLab-OpenDoge-Getup(平地倒地自恢复)、LainLab-OpenDoge-Handstand(倒立)、LainLab-OpenDoge-Jump(跳跃)
  • 单策略多技能:LainLab-OpenDoge-Skills-Flat,一个网络在 walk / handstand / getup / jump 之间切换,57 维观测、12 维动作

仓库同时保留 Unitree 机型的资产与任务(A2、As2、Go2、G1、G1-23Dof、H1_2、H2、R1,以及 Go2 的 DreamWaQ / AMP-DreamWaQ / CTS / TS Teacher 技能迁移),用于对照 mjlab 上游能力、验证共享 MDP 的通用性。新增机型沿用同一条路径:资产 → profile/任务 → 验证 → 部署适配。

任务 ID 前缀:LainLab- 是历史命名,代码骨架沿用自 LainLab 项目,为兼容已经发布的 baseline 与 checkpoint 暂未改名。改成 OpenDoge- 前缀需要同步调整 src/tasks/robots/opendoge/、baseline 目录与测试,留待单独提交。

项目定位

维度 上游 mjlab unitree_rl_mjlab 参考项目 本仓库 isaacgym 分支 OpenDoge Train(mjlab)
角色 通用 MuJoCo/Warp 仿真与强化学习框架 以 Unitree 机型为中心的训练与实机部署项目 HIMLoco 衍生的 Isaac Gym 训练栈 OpenDoge 在 mjlab 上的训练、验证与导出层
机器人范围 通用框架能力 Unitree 机型 OpenDoge / A1 / Go1 以 OpenDoge 为主,Unitree 机型作为对照保留
框架维护 上游维护仿真、manager、MDP、runner 与 viewer 项目自带机型任务与部署流程 自带 legged_gym 与 PPO/AMP 实现 固定依赖 mjlab==1.6.0,用薄适配层避免复制框架
任务组织 通用任务与扩展机制 速度跟踪、动作模仿、实机控制 legged_gym/envs 下的机型配置 机器人 profile + 独立资产 + 可注册任务表达差异
可验证性 框架级能力 训练/部署工作流 MuJoCo Sim2Sim + 真机 资产、配置、观测/奖励契约与 CPU smoke 测试,已知限制公开记录
动作数据 通用工具与接口 面向 Unitree 工作流 AMP / 关键点数据集 本地 CSV → NPZ,按目标机器人校验关节列与自由度,不依赖 W&B

目录结构

OpenDoge_train/
├── src/                 # 源码根(Python 包为 src.*,发行包名 lainlab)
│   ├── assets/          # 各机器人 MJCF、URDF、网格与示例动作
│   ├── tasks/           # 任务类型与机器人组成
│   │   ├── velocity/    # 通用速度跟踪 MDP
│   │   ├── tracking/    # 通用动作跟踪 MDP
│   │   ├── getup/       # 起身任务类型
│   │   ├── handstand/   # 倒立任务类型
│   │   ├── jump/        # 跳跃任务类型
│   │   ├── multiskill/  # 单策略多技能任务类型
│   │   ├── amp/         # 预留通用 AMP core
│   │   └── robots/      # 机器人组成、profile 与任务注册
│   ├── toolchain/       # 离线工具链:标定 / 采集 / BC / DAgger / 导出
│   ├── workbench/       # 独立工作台(HTTP 服务与查看器)
│   ├── baseline/        # 版本化基线策略(随包分发)
│   ├── baselines.py     # 基线发现
│   ├── cli.py           # train / play / list-envs 入口
│   └── viewer.py        # Viser 回放扩展
├── tests/               # 兼容性测试
└── docs/                # 设计、验收与工作台文档

分层原则:

  • mjlab 负责仿真、manager、通用 MDP、runner 和 viewer。
  • src.assets 负责 OpenDoge 等机器人的模型描述与动作数据。
  • src.tasks 只表达机器人名称、接触点、动作缩放等差异。
  • CLI 先注册本仓库任务,再复用 mjlab 1.6 的训练与回放实现。

安装

需要 Python 3.10–3.13;训练需要 NVIDIA GPU,CPU 只适合配置检查与 smoke 测试。

uv sync --extra cu128

仅做 CPU 配置检查时:

uv sync --extra cpu

使用

# 查看任务
uv run list-envs

# 训练 OpenDoge
uv run train LainLab-OpenDoge-Flat --env.scene.num-envs 4096

# 直接播放内置 baseline
uv run baseline-play opendoge flat --device cuda:0 --num-envs 1
uv run baseline-play opendoge rough --device cuda:0 --num-envs 1 --no-terminations True
uv run baseline-play opendoge getup --device cuda:0 --num-envs 1

# 用随机动作做配置冒烟测试
uv run play LainLab-OpenDoge-Flat --agent random --num-envs 1

# 单策略多技能:walk / handstand / getup / jump
uv run train LainLab-OpenDoge-Handstand --env.scene.num-envs 4096
uv run train LainLab-OpenDoge-Jump --env.scene.num-envs 4096
uv run train LainLab-OpenDoge-Skills-Flat --env.scene.num-envs 4096
uv run baseline-play opendoge skills --num-envs 16
uv run opendoge-calibrate --write          # 倒立/跳跃目标标定
uv run opendoge-eval LainLab-OpenDoge-Skills-Flat --checkpoint <ckpt>
uv run opendoge-eval LainLab-OpenDoge-Skills-Flat --checkpoint <ckpt> --transitions

# 回放本地策略
uv run play LainLab-OpenDoge-Flat --checkpoint-file logs/.../model_1000.pt

# 导出 ONNX
uv run opendoge-export LainLab-OpenDoge-Flat --checkpoint <ckpt> --output policy.onnx

# 把 G1 CSV 动作转换为 mjlab 跟踪格式
uv run csv-to-npz --input-file src/assets/motions/g1/dance1_subject2.csv \
  --output-name dance1-subject2 --robot g1

内置 baseline 位于 src/baseline/<robot>/<terrain>/,每个条目带 baseline.json 清单和 model.pt / policy.onnx,说明见 src/baseline/README.md。

续训提示:Unitree 速度任务的实验名已由 <robot>_velocity 改为 <task_group>_<robot>_<terrain>_velocity(Flat 与 Rough 不再共用同一目录)。 续训改名前的实验需显式指定旧目录名,例如: uv run train Unitree-Go2-Flat --agent.resume --agent.load-run go2_velocity。

独立工作台

# 在仓库根目录启动(webui / src-webui 兼容入口仍可用)
uv run --extra workbench src-workbench

浏览器打开 http://127.0.0.1:7860。工作台按固定顺序组织:

  1. 项目内 URDF/MJCF 查看器:加载项目模型,或导入包含网格的目录;调节关节,显示坐标系、关节轴、碰撞体和惯量。
  2. 策略选择:Velocity / Tracking 互斥展开;Tracking 内置 GMR 的 LAFAN1 BVH → G1 29-DoF 重定向,从统一动作资产库选择输入,也支持已有 GMR PKL → NPZ;实时查看重定向进度、人体骨架与 G1 姿态,完成后通过时间轴回放。
  3. 算法选择:Velocity / Tracking 分组;Go2 Skill 单独提供 DreamWaQ、AMP-DreamWaQ、CTS、TS Teacher 和动作模板。
  4. 训练:设置迭代轮次、保存间隔、并行环境数;查看实时日志、停止任务,在内嵌 TensorBoard 中查看奖励和训练指标。
  5. 导出与验证:浏览、选择、下载 checkpoint/ONNX,手动导出 ONNX,启动 CPU 单环境 Viser checkpoint 回放。

工作台源码在 src/workbench/,不改变 tasks/ 中的训练算法与环境。查看器由项目自行实现:MuJoCo 解析模型,精简 Three.js 页面负责绘制;仅参考 robot_viewer 的功能思路,不复制其整套应用。查看器与 GMR 子集都在项目内运行。

模型导入用于检查;训练仍使用已注册任务的机器人配置。GMR 当前限 LAFAN1 骨架和 G1 29-DoF,23-DoF 任务需使用对应 NPZ。Viser 当前回放 checkpoint,ONNX 是独立导出产物。

动作集中在 src/assets/motions/。页面支持导入和选择文件,转换输出自动命名、入库,并选为训练动作。

前端开发、模块边界和验证说明见 工作台文档。

在线仿真器

浏览器版 MuJoCo + ONNX 策略演示已迁出本仓库,独立部署在 https://opendogerobotics.github.io/(源码:OpenDogeRobotics.github.io)。训练、评估和工作台都不依赖任何前端资源。

动作数据

csv-to-npz 直接在本地生成 NPZ,不需要 Weights & Biases。--output-name 只传文件名时,输出会保存到输入 CSV 的同一目录;也可传入完整路径。G1-23DoF 动作使用 --robot g1_23dof。

任务命名统一使用分类前缀:Unitree-<Robot>-Flat / Unitree-<Robot>-Rough 和 LainLab-<Robot>-Flat / LainLab-<Robot>-Rough。速度任务支持 A2、As2、Go2、G1、G1-23Dof、H1_2、H2、R1 和 OpenDoge;动作跟踪任务支持 G1 与 G1-23Dof。OpenDoge 的 Flat、Rough、Getup、Handstand、Jump 与 LainLab-OpenDoge-Skills-Flat 均已注册,单策略多技能的完整设计与验收见 docs/opendoge_multiskill.md。

运行验证与已知限制

截至 2026-09-09,项目的 CPU 单环境 smoke 验证覆盖了已注册任务的环境构建、reset 与一步零动作执行,并检查了观测和奖励是否为有限值。所有 Go2 任务(含 Go2 skill)、基础速度任务、Cartpole、Yam 操作任务和 29-DoF G1 跟踪任务均已通过;OpenDoge 的各任务另有 opendoge-eval 的可复现指标。

  • 跟踪任务必须提供与机器人自由度匹配的 NPZ 动作文件。命令行训练/回放时请设置 motion_file;仓库本地的 29-DoF G1 示例不能用于 G1-23Dof 跟踪任务。
  • Unitree-H2-Rough 和 Unitree-R1-Rough 当前的 MuJoCo-Warp 接触缓冲容量不足:分别至少需要 nconmax=80 和 nconmax=141。提高该配置值后,两项任务可完成 smoke;在修复默认配置前请勿直接用默认设置启动训练。
  • Smoke 只验证初始化和单步数值稳定性,不代表策略已经收敛,也不替代长时间训练、checkpoint 回放或真机验收。

开发

make format
make check

机器人差异集中在 src/tasks/robots/ 下的机器人包,通用 MDP 位于 src/tasks/velocity/core.py、tracking/core.py 等任务类型包。新增同类机器人时,通常在对应机器人包中增加 profile 和注册函数,无需复制整套 MDP。

部署与导出

本分支不包含实机部署源码:训练与验证通过 play 的 Viser 回放完成,导出的 ONNX 策略(opendoge-export)交给实机侧工程使用,例如 OpenDoge 固件仓库 OpenDoge_firmware。Isaac Gym 版本中的 deploy/deploy_mujoco/(配置驱动的 MuJoCo 验证与诊断)仅存在于 isaacgym 分支。

上游与许可

本分支是 mjlab 1.6.0 的下游项目,项目组织参考 unitree_rl_mjlab,代码骨架与任务前缀沿袭自 LainLab。代码使用 Apache-2.0 许可;OpenDoge 资产由 OpenDogeRobotics 维护,第三方组件与机器人描述保留各自许可。


OpenDoge Train (mjlab)

Hardware

OpenDoge is an open-source quadruped robot from OpenDogeRobotics. This repository, OpenDoge_train, owns its reinforcement-learning training, simulation validation, and policy export.

The repository carries two parallel branches with different stacks and non-interchangeable commands:

  • mjlab (default; described by this README): the current mainline, built on mjlab 1.6.0 (MuJoCo-Warp), with OpenDoge velocity/get-up/handstand/jump/multi-skill tasks, Viser playback, a standalone workbench, and a separate online browser simulator.
  • isaacgym: the original Isaac Gym stack (Preview 4 + legged_gym + HIMLoco-derived PPO/AMP), including MuJoCo Sim2Sim validation and ONNX export. Kept for reproducing earlier results and for comparison.
git checkout isaacgym   # the Isaac Gym version
git checkout mjlab      # back to the mainline

Branch comparison

Branch Stack Contents Role
mjlab (default) Python 3.10–3.13 + mjlab 1.6.0 + MuJoCo-Warp + RSL-RL + PyTorch Task registration for OpenDoge and Unitree robots, Viser playback, workbench, multi-skill policies, ONNX export Current development
isaacgym Python 3.8 + Isaac Gym Preview 4 + legged_gym + HIMLoco rsl_rl OpenDoge/A1/Go1 configs, MuJoCo Sim2Sim, ONNX export Historical version, maintenance fixes

Why mjlab

  • No Isaac Gym dependency. mjlab is pure Python + MuJoCo-Warp, so a single CPU environment can exercise environment construction, reset, and one action step; development machines do not need a GPU or an Isaac Gym install.
  • Assets and tasks in one registry. Robot assets live in src/assets/, task differences in src/tasks/robots/, and training, playback, and the workbench share the same task registry.
  • Pinned framework, thin adapter. mjlab==1.6.0 is a fixed dependency; this repository only owns robot assets, task differences, validation contracts, and deployment adapters.
  • One CLI for training, validation, and export. train, play, baseline-play, opendoge-eval, and opendoge-export all ship in the repository.

Robots and tasks

OpenDoge is the primary robot of this repository:

  • Velocity tracking: LainLab-OpenDoge-Flat, LainLab-OpenDoge-Rough
  • Single-skill tasks: LainLab-OpenDoge-Getup (fall recovery on flat ground), LainLab-OpenDoge-Handstand, LainLab-OpenDoge-Jump
  • Single-policy multi-skill: LainLab-OpenDoge-Skills-Flat, one network switching between walk / handstand / getup / jump with a 57-D observation and 12-D action

Unitree assets and tasks (A2, As2, Go2, G1, G1-23Dof, H1_2, H2, R1, plus the Go2 DreamWaQ / AMP-DreamWaQ / CTS / TS Teacher skill migrations) remain as reference and as coverage for the shared MDPs. New robots follow the same path: asset → profile/task → validation → deployment adapter.

Task-ID prefix: LainLab- is a historical name inherited from the LainLab codebase. It is kept for compatibility with already-published baselines and checkpoints; renaming to an OpenDoge- prefix requires updating src/tasks/robots/opendoge/, the baseline directories, and tests together, so it is left for a separate commit.

Positioning

Dimension Upstream mjlab unitree_rl_mjlab reference isaacgym branch OpenDoge Train (mjlab)
Role General MuJoCo/Warp simulation and RL framework Unitree-centered training and real-robot deployment HIMLoco-derived Isaac Gym stack Training, validation, and export layer for OpenDoge on mjlab
Robot scope General framework capabilities Unitree robots OpenDoge / A1 / Go1 OpenDoge first; Unitree robots kept for comparison
Framework maintenance Upstream owns simulation, managers, MDPs, runners, viewers Carries robot tasks and deployment workflows Carries legged_gym and PPO/AMP implementations Pins mjlab==1.6.0; thin adapters avoid copying the framework
Task organization General tasks and extension points Velocity tracking, motion imitation, real-robot control Per-robot configs under legged_gym/envs Robot profiles, isolated assets, and registrable tasks express differences
Verifiability Framework-level Training/deployment workflow MuJoCo Sim2Sim and real hardware Asset, config, observation/reward contract, and CPU smoke tests; known limits documented
Motion data General tooling Unitree motion workflows AMP and keypoint datasets Local CSV → NPZ with joint-column and DoF validation per robot, no W&B

Layout

OpenDoge_train/
├── src/                 # Source root (Python package: src.*, distribution name: lainlab)
│   ├── assets/          # Per-robot MJCF/URDF files, meshes, example motions
│   ├── tasks/           # Task types and robot composition
│   │   ├── velocity/    # Shared velocity-tracking MDP
│   │   ├── tracking/    # Shared motion-tracking MDP
│   │   ├── getup/       # Get-up task type
│   │   ├── handstand/   # Handstand task type
│   │   ├── jump/        # Jump task type
│   │   ├── multiskill/  # Single-policy multi-skill task type
│   │   ├── amp/         # Reserved generic AMP core
│   │   └── robots/      # Robot composition, profiles, registration
│   ├── toolchain/       # Offline toolchain: calibrate / collect / BC / DAgger / export
│   ├── workbench/       # Standalone workbench (HTTP server + viewer)
│   ├── baseline/        # Versioned baseline policies (shipped in the wheel)
│   ├── baselines.py     # Baseline discovery
│   ├── cli.py           # train / play / list-envs entry points
│   └── viewer.py        # Viser playback extensions
├── tests/               # Compatibility tests
└── docs/                # Design, acceptance, and workbench docs

Layering principles:

  • mjlab provides simulation, managers, common MDPs, runners, and viewers.
  • src.assets owns the OpenDoge and reference robot descriptions plus motion data.
  • src.tasks expresses only task differences such as robot names, contacts, and action scales.
  • The CLI registers this repository's tasks before reusing mjlab 1.6 training and playback.

Installation

Python 3.10–3.13 is required. Training needs an NVIDIA GPU; CPU is only for configuration checks and smoke tests.

uv sync --extra cu128

For CPU-only configuration checks:

uv sync --extra cpu

Usage

# List tasks
uv run list-envs

# Train OpenDoge
uv run train LainLab-OpenDoge-Flat --env.scene.num-envs 4096

# Play bundled baselines directly in Viser
uv run baseline-play opendoge flat --device cuda:0 --num-envs 1
uv run baseline-play opendoge rough --device cuda:0 --num-envs 1 --no-terminations True
uv run baseline-play opendoge getup --device cuda:0 --num-envs 1

# Random-action configuration smoke test
uv run play LainLab-OpenDoge-Flat --agent random --num-envs 1

# Single-policy multi-skill: walk / handstand / getup / jump
uv run train LainLab-OpenDoge-Handstand --env.scene.num-envs 4096
uv run train LainLab-OpenDoge-Jump --env.scene.num-envs 4096
uv run train LainLab-OpenDoge-Skills-Flat --env.scene.num-envs 4096
uv run baseline-play opendoge skills --num-envs 16
uv run opendoge-calibrate --write          # handstand / jump target calibration
uv run opendoge-eval LainLab-OpenDoge-Skills-Flat --checkpoint <ckpt>
uv run opendoge-eval LainLab-OpenDoge-Skills-Flat --checkpoint <ckpt> --transitions

# Replay a local policy
uv run play LainLab-OpenDoge-Flat --checkpoint-file logs/.../model_1000.pt

# Export ONNX
uv run opendoge-export LainLab-OpenDoge-Flat --checkpoint <ckpt> --output policy.onnx

# Convert a G1 CSV motion into mjlab tracking format
uv run csv-to-npz --input-file src/assets/motions/g1/dance1_subject2.csv \
  --output-name dance1-subject2 --robot g1

Bundled baselines live in src/baseline/<robot>/<terrain>/; each entry has a baseline.json manifest plus model.pt / policy.onnx. See src/baseline/README.md.

Resuming runs: Unitree velocity experiment names changed from <robot>_velocity to <task_group>_<robot>_<terrain>_velocity (Flat and Rough no longer share a directory). To resume a pre-rename run, point at its old experiment directory explicitly, e.g. uv run train Unitree-Go2-Flat --agent.resume --agent.load-run go2_velocity.

Standalone workbench

# Run from the repository root (webui / src-webui remain as compatible entry points)
uv run --extra workbench src-workbench

Open http://127.0.0.1:7860. The workbench is organized as a fixed five-step flow:

  1. In-project URDF/MJCF viewer: load project models or import a directory with meshes; move joints, show frames, joint axes, collision bodies, and inertias.
  2. Policy selection: Velocity and Tracking expand exclusively; Tracking bundles GMR's LAFAN1 BVH → G1 29-DoF retargeting from the shared motion library, also accepts existing GMR PKL → NPZ, and shows retargeting progress, the human skeleton, and the G1 pose before timeline playback.
  3. Algorithm selection: Velocity / Tracking groups; Go2 skills expose DreamWaQ, AMP-DreamWaQ, CTS, TS Teacher, and motion templates separately.
  4. Training: set iterations, save interval, and parallel environments; watch live logs, stop jobs, and inspect rewards and metrics in the embedded TensorBoard.
  5. Export and validation: browse, select, and download checkpoints/ONNX, export ONNX manually, and launch CPU single-environment Viser checkpoint playback.

The workbench source lives in src/workbench/ and does not change training algorithms or environments under tasks/. The viewer is implemented in-project: MuJoCo parses the model and a small Three.js page draws it, following robot_viewer's feature outline without copying the application. The viewer and the GMR subset both run inside the project.

Model import is for inspection; training still uses the registered task's robot configuration. GMR currently supports the LAFAN1 skeleton and G1 29-DoF, so 23-DoF tasks need their own NPZ. Viser plays checkpoints; ONNX is a separate export artifact.

Motions live in src/assets/motions/. The page imports and selects files, names and stores converted outputs, and can pick them as training motions.

Front-end development, module boundaries, and validation notes are in the workbench documentation.

Online simulator

The browser MuJoCo + ONNX demo now lives outside this repository at https://opendogerobotics.github.io/ (source: OpenDogeRobotics.github.io). Training, evaluation, and the workbench have no front-end dependency.

Motion data

csv-to-npz creates the NPZ locally and does not require Weights & Biases. If --output-name is only a filename, the output is written next to the input CSV; an absolute or relative output path may also be used. Use --robot g1_23dof for G1-23DoF motions.

Tasks use a category prefix: Unitree-<Robot>-Flat / Unitree-<Robot>-Rough and LainLab-<Robot>-Flat / LainLab-<Robot>-Rough. Velocity tasks cover A2, As2, Go2, G1, G1-23Dof, H1_2, H2, R1, and OpenDoge; motion tracking covers G1 and G1-23Dof. OpenDoge registers Flat, Rough, Getup, Handstand, Jump, and LainLab-OpenDoge-Skills-Flat; the multi-skill design and acceptance record is in docs/opendoge_multiskill.md.

Validation status and known limitations

As of 2026-09-09, CPU single-environment smoke checks cover environment construction, reset, and one zero-action step for registered tasks, checking that observations and rewards stay finite. All Go2 tasks (including Go2 skills), baseline velocity tasks, Cartpole, Yam manipulation tasks, and 29-DoF G1 tracking tasks passed; OpenDoge tasks additionally have reproducible metrics from opendoge-eval.

  • Tracking tasks need an NPZ motion file whose degrees of freedom match the robot. Set motion_file for training or playback. The local 29-DoF G1 example cannot be used for G1-23Dof tracking.
  • The default MuJoCo-Warp contact-buffer capacity is insufficient for Unitree-H2-Rough and Unitree-R1-Rough: they need at least nconmax=80 and nconmax=141, respectively. Both pass smoke checks after that adjustment; do not start training with their defaults until the configuration is fixed.
  • Smoke tests validate initialization and single-step numerical stability only. They do not demonstrate convergence and do not replace long training, checkpoint playback, or real-robot validation.

Development

make format
make check

Robot differences are centralized under src/tasks/robots/. Shared MDPs live in task-type packages such as velocity/core.py and tracking/core.py. Adding a similar robot is typically a profile plus a registration in its robot package, without copying the shared MDP implementation.

Deployment and export

This branch does not ship real-robot deployment code: training and validation go through the play Viser viewer, and exported ONNX policies (opendoge-export) are handed to the robot-side project, e.g. the OpenDoge firmware repository OpenDoge_firmware. The Isaac Gym version's deploy/deploy_mujoco/ (config-driven MuJoCo validation and diagnostics) exists only on the isaacgym branch.

Upstream and license

This branch is a downstream project of mjlab 1.6.0, organized after unitree_rl_mjlab; the code skeleton and task prefix are inherited from LainLab. The code is Apache-2.0 licensed; OpenDoge assets are maintained by OpenDogeRobotics, and third-party components and robot descriptions keep their own licenses.

About

No description, website, or topics provided.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages