Getting Started

The project consists of two sub-projects, Simulation Platform and Imitation Learning Platform, both running in containers. This page covers repository setup, requirements, and the overall workflow; follow the robot-specific documentation for detailed steps.

Get the code

After cloning the repository, pull the LFS assets (USD models, textures, etc.) and initialize the lerobot submodule.

# (Optional) Configure a proxy if GitHub is slow; adjust the port to your local proxy
git config --global http.proxy  http://127.0.0.1:7897
git config --global https.proxy http://127.0.0.1:7897
export GIT_LFS_PROXY="http://127.0.0.1:7897"   # route LFS through the proxy

# 1. Clone without LFS files and submodules for a faster checkout
GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/LQX-Code-SH/Thinker-IL-LAB.git
cd Thinker-IL-LAB

# 2. Pull LFS assets
git lfs pull

# 3. Initialize and update submodules (lerobot)
git submodule update --init

LFS assets are ubt_sim simulation resources (*.usd, *.urdf, textures, etc.). If you only need real-robot data conversion, training, or deployment, you can skip LFS. The lerobot submodule is required; ubt_sim is only needed for simulation data collection.

Requirements

Dependency	Requirement	Notes
OS	Linux (Ubuntu 22.04 recommended)	Both sub-projects run in containers
GPU	NVIDIA GPU (RTX 3090 or better)	Isaac Sim rendering and policy training/inference
Docker	NVIDIA Container Toolkit installed	GPU access inside containers
Git LFS	Installed	Manages USD models, textures, and other large files
Disk	Sufficient space (100 GB+ recommended)	LFS assets, datasets, and checkpoints can be large

End-to-end workflow

In one sentence: collect HDF5 data -> convert to a LeRobot dataset -> train in the container -> evaluate offline -> deploy (simulation or real robot).

Simulation collection (ubt_sim) ───┐
                                   ├─► data conversion ─► training ─► evaluation ─► sim validation ─► real deployment
Real-robot collection (teleop) ────┘

 

There are two ways to get familiar with the project: explore the platforms first (ubt_sim simulation collection / ubt_IL imitation learning), or jump directly into an end-to-end robot workflow (Walker TienKung Pro / Walker TienKung Dex / Walker S2 EDU / Walker C1 EDU) and learn by reproducing it.

Browse by platform

Start from the platform level to understand the shared architecture, commands, and workflows:

Platform	Overview
Simulation Platform	Simulation Platform Guide
Imitation Learning Platform	Imitation Learning Platform Guide

Browse by robot

Follow a complete documentation track for each robot:

Robot	Overview	End-to-end workflow
Walker TienKung Pro	Walker TienKung Pro Overview	Walker TienKung Pro Simulation Workflow / Walker TienKung Pro Real-Robot Workflow
Walker TienKung Dex	Walker TienKung Dex Overview	Walker TienKung Dex Simulation Workflow / Walker TienKung Dex Real-Robot Workflow
Walker S2 EDU	Walker S2 EDU Overview	Walker S2 EDU Simulation Workflow / Walker S2 EDU Real-Robot Workflow
Walker C1 EDU	Walker C1 EDU Overview	Walker C1 EDU Simulation Workflow / Walker C1 EDU Real-Robot Workflow

Optional: download sample data and models

Sample datasets and trained ACT models are available on HuggingFace (datasets / models). Download them from the website or manage them with the bundled hf_manager.py tool:

cd ubt_IL/scripts/convert/common
export HF_ENDPOINT=https://hf-mirror.com          # (Optional) mirror for faster access in China

# Download a dataset or model
python hf_manager.py pull <dataset-or-model-path>
# Examples
python hf_manager.py pull qingxiangliu/tienkung_pick_up_merged       # real-robot dataset
python hf_manager.py pull qingxiangliu/tienkung_pick_up_act          # real-robot ACT policy

Datasets

Walker TienKung Pro

Repository	Size	Format	Notes
tienkung_sim_pick_place ★	1.1 MB	LeRobot	Simulated apple pick-and-place
tienkung_pick_up_merged ★	1.6 GB	LeRobot	Merged real-robot pick-up dataset
tienkung_pick_up_the_apple_all	426.9 MB	LeRobot	Real-robot apple pick-up
tienkung_pick_up_the_bottle_all	481.0 MB	LeRobot	Real-robot bottle pick-up
tienkung_pick_up_the_red_bottle_all	685.2 MB	LeRobot	Real-robot red-bottle pick-up
tienkung_real_merged	379.3 MB	LeRobot	Real-robot bottle pick-up (Guochuang)

Walker S2 EDU

Repository	Size	Format	Notes
Walker_S2_sim_10_2RGB ★	119.2 MB	LeRobot	Simulated part picking
Walker-s2-pick-part-sim	4.3 GB	HDF5	Simulated part picking
Walker_S2_real_10_2RGB	157.5 MB	LeRobot	Real-robot multi-part picking
Wlaker_Pick_part_real_10d_2RGB ★	24.2 MB	LeRobot	Real-robot part picking
walker-s2-pick-part-real-hdf5	21.6 GB	HDF5	Real-robot multi-part picking
walker_pick_part_real_hdf5	9.7 GB	HDF5	Real-robot part picking
walker_pick_sort	11.5 GB	HDF5	Real-robot part sorting

Models (ACT)

Repository	Size	Notes
tienkung_sim_pick_place_right13_act	591 MB	Walker TienKung Pro simulated ACT policy
tienkung_pick_up_act	591 MB	Walker TienKung Pro real-robot ACT policy
Walker_S2_sim_10_2RGB_act	1.9 GB	Walker S2 EDU simulated ACT policy
walker_pick_part_real_10d_2RGB_act	1.7 GB	Walker S2 EDU real-robot ACT policy

Datasets are stored in ubt_IL/dataset/ and models in ubt_IL/model/. See HF Data & Models for details.