Embodied Lab

English

Embodied Lab is an offline-first research stack for a JAKA Mini2 arm and Inspire RH56DFX hand. Current maintained functions are:

Quest 3 input parsing and safe target generation;
MuJoCo simulation and replay;
explicit, separately authorized JAKA/RH56 hardware adapters;
review-first RGB-D/robot episode collection and dataset preparation;
ACT/LeRobot and π0.5/OpenPI training infrastructure.

The current arm path is:

Quest HTS + CTRL
  -> validate and order input
  -> release-before-press reference capture
  -> map, filter, continuation IK, and feasibility checks
  -> immutable AcceptedArmTarget
  -> MuJoCo adapter or JAKA joint adapter

The physical adapter never follows MuJoCo qpos, remaps a target, or solves IK. Native joint teleoperation makes zero JAKA kine_inverse calls.

Safety boundary

Tests, replay, simulation, doctor, and --help are offline actions. They do not connect to JAKA, RH56DFX, Quest, RealSense, or any actuator. Physical operation requires an explicit operator procedure and the safety rules in docs/safety/REAL_HARDWARE_SAFETY.md.

Quick start

python3 -m venv .venv
.venv/bin/python -m pip install -e ".[dev]"
.venv/bin/embodied-lab doctor
.venv/bin/embodied-lab sim smoke
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m pytest -q -p no:cacheprovider

Start with the documentation index. The maintained entry points are grouped in capabilities, dataset, ACT training, and π0.5 training.

Repository layout

Path	Current role
`src/`	Reusable input, control, safety, RH56, camera, and dataset packages
`native/`	Offline-buildable JAKA and teleoperation workers
`tools/`	Current operator and developer entrypoints
`scripts/`	Thin wrappers for current simulation, collection, and training workflows
`configs/`	Current runtime and training configuration
`training/`	Project-owned ACT and π0.5 training boundaries
`assets/`	Current MuJoCo and robot assets
`docs/`	Current functional documentation
`third_party/`	Pinned submodules and vendor/reference assets

中文

Embodied Lab 是面向 JAKA Mini2 机械臂和 Inspire RH56DFX 灵巧手的离线优先研究栈。当前维护的功能包括：

Quest 3 输入解析和安全 target 生成；
MuJoCo 仿真与 replay；
必须单独授权的 JAKA/RH56 真机适配器；
先人工审核的 RGB-D/机器人 episode 采集和数据准备；
ACT/LeRobot 与 π0.5/OpenPI 训练基础设施。

当前机械臂管线为：

Quest HTS + CTRL
  -> 输入校验和排序
  -> release-before-press reference capture
  -> 映射、滤波、continuation IK 和 feasibility 检查
  -> 不可变 AcceptedArmTarget
  -> MuJoCo adapter 或 JAKA joint adapter

物理适配器不会跟随 MuJoCo qpos、重新映射 target 或重新求 IK。原生 joint teleoperation 不会调用 JAKA kine_inverse。

安全边界

测试、replay、仿真、doctor 和 --help 都是离线操作，不会连接 JAKA、RH56DFX、Quest、RealSense 或任何 actuator。真机操作必须经过明确的 operator procedure，并遵守 docs/safety/REAL_HARDWARE_SAFETY.md。

快速开始

python3 -m venv .venv
.venv/bin/python -m pip install -e ".[dev]"
.venv/bin/embodied-lab doctor
.venv/bin/embodied-lab sim smoke
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m pytest -q -p no:cacheprovider

请从文档索引开始。当前入口集中在能力说明、数据集、ACT 训练和 π0.5 训练。

仓库布局

路径	当前职责
`src/`	可复用的输入、控制、安全、RH56、相机和数据集包
`native/`	可离线构建的 JAKA 和 teleoperation worker
`tools/`	当前 operator/developer 入口
`scripts/`	当前仿真、采集和训练流程的薄 wrapper
`configs/`	当前 runtime 和 training 配置
`training/`	项目维护的 ACT 和 π0.5 training boundary
`assets/`	当前 MuJoCo 和机器人资产
`docs/`	当前功能文档
`third_party/`	固定版本的 submodule 和 vendor/reference 资产