Huy-Quang-Dao/instantTrack

选用指南

AI 根据 README 整理 · 2026/10/29 23:46:17

适用场景

适用于在仿真环境中训练 Unitree G1 机器人的全身运动跟踪策略，基于 LAFAN1 数据集，可复现论文中的实验结果。

查看 README 依据

# InstantTrack **Whole-body motion tracking for the Unitree G1 in 15 minutes on one GPU.** InstantTrack trains a single policy for all 40 LAFAN1 clips (2.45 h of motion) to an evaluation normalized return of 0.82 with **0/40 failed clips in 0.25 h on one RTX 5090**. On the same machine, simulator, MDP and evaluator, our reimplementation of SuperMimic needs **1.50 h**, and SuperMimic with the policy EMA we add needs **0.55 h**. InstantTrack is also more accurate and makes cleaner foot contact. InstantTrack builds on SuperMimic (one differentiable world model, backpropagation through imagined

- Without the contact reward, InstantTrack beats SuperMimic + EMA on every metric we measure. - **Scope.** Only the Unitree G1 in simulation, on LAFAN1 (40 clips). Evaluation has no domain randomization and no pushes. There is no sim-to-real experiment and no larger motion set. ## Setup **Requirements:** - Linux x86_64. - An NVIDIA GPU with a driver for CUDA 12.8 (tested: RTX 5090, driver 580; CUDA 12.8 needs ≥ 570). - [uv](https://docs.astral.sh/uv/getting-started/installation/), `git`, `curl`. - `ffmpeg`, only for rendering videos.

目标用户

面向机器人运动控制领域的研究人员和开发者，需要熟悉 MuJoCo 和 PyTorch。

查看 README 依据

# InstantTrack **Whole-body motion tracking for the Unitree G1 in 15 minutes on one GPU.** InstantTrack trains a single policy for all 40 LAFAN1 clips (2.45 h of motion) to an evaluation normalized return of 0.82 with **0/40 failed clips in 0.25 h on one RTX 5090**. On the same machine, simulator, MDP and evaluator, our reimplementation of SuperMimic needs **1.50 h**, and SuperMimic with the policy EMA we add needs **0.55 h**. InstantTrack is also more accurate and makes cleaner foot contact. InstantTrack builds on SuperMimic (one differentiable world model, backpropagation through imagined

rollouts) and adds InstantMimic-style GPU work on the training loop, a few algorithmic changes and a contact reward. It runs on [mjlab](https://github.com/mujocolab/mjlab) / MuJoCo Warp and is installed with [uv](https://docs.astral.sh/uv/). **Project page:** <https://huy-quang-dao.github.io/instantTrack/>  *Both policies trained for 15 minutes (0.25 h) on one RTX 5090. Left: SuperMimic + EMA loses

使用方式

通过命令行使用，提供训练、评估、渲染等脚本，支持自定义参数。

查看 README 依据

| `instanttrack_track` | all InstantTrack changes except the contact reward | best normalized return and pose accuracy | | `instanttrack` | `--pi_ema 0.995 --no_sync --fused_adam --fused_rot --compile_more --lr 4e-4 --wm_batches 5 --wm_hidden 1024,1024,1024 --fail_reset 0.3 --contact_reward 0.25` | **main configuration** | ```bash # train InstantTrack for 0.75 h (seed 0), then evaluate all its checkpoints scripts/train.sh instanttrack it_s0 --seed 0 --max_hours 0.75 # the baselines scripts/train.sh supermimic_ema sm_ema_s0 --seed 0 --max_hours 0.75

uv run python tools/eval_batch.py --run_dir logs/it_s0 # every checkpoint -> logs/it_s0/eval.csv uv run python tools/compare.py logs/sm_ema_s0 logs/it_s0 --out results/compare # time-to-target table + curves uv run python tools/eval_metrics.py logs/it_s0/ckpt_017324.pt --out results/metrics # pose / contact metrics uv run python -m instanttrack.evaluate --run_dir logs/it_s0 --ckpt logs/it_s0/ckpt_017324.pt --record traj.npz uv run python tools/render.py traj.npz video.mp4 --clips fallAndGetUp1_subject1 jumps1_subject1 --seconds 12 ```

部署要求

需要 Linux x86_64 系统、NVIDIA GPU（支持 CUDA 12.8，驱动版本 ≥570）、uv、git、curl，以及约 10 GB 磁盘空间。

查看 README 依据

- Without the contact reward, InstantTrack beats SuperMimic + EMA on every metric we measure. - **Scope.** Only the Unitree G1 in simulation, on LAFAN1 (40 clips). Evaluation has no domain randomization and no pushes. There is no sim-to-real experiment and no larger motion set. ## Setup **Requirements:** - Linux x86_64. - An NVIDIA GPU with a driver for CUDA 12.8 (tested: RTX 5090, driver 580; CUDA 12.8 needs ≥ 570). - [uv](https://docs.astral.sh/uv/getting-started/installation/), `git`, `curl`. - `ffmpeg`, only for rendering videos.

- About 10 GB of disk for the environment and data, plus about 2 GB per 0.75 h run for checkpoints. ```bash git clone https://github.com/Huy-Quang-Dao/instantTrack.git cd instantTrack uv sync # creates .venv: mjlab, MuJoCo Warp, Warp, torch 2.9.0 (CUDA 12.8) tools/download_data.sh # LAFAN1 G1 retargets (40 clips) + G1 description -> data/, builds data/g1_description/g1_29dof.xml uv run python tools/test_env.py # sanity check: FK vs simulator (~1e-6), state round-trip, reference playback ```

已知限制

仅支持仿真环境中的 Unitree G1 机器人，使用 LAFAN1 数据集（40 个片段），无 sim-to-real 实验，且运动数据不随仓库分发。

查看 README 依据

- Without the contact reward, InstantTrack beats SuperMimic + EMA on every metric we measure. - **Scope.** Only the Unitree G1 in simulation, on LAFAN1 (40 clips). Evaluation has no domain randomization and no pushes. There is no sim-to-real experiment and no larger motion set. ## Setup **Requirements:** - Linux x86_64. - An NVIDIA GPU with a driver for CUDA 12.8 (tested: RTX 5090, driver 580; CUDA 12.8 needs ≥ 570). - [uv](https://docs.astral.sh/uv/getting-started/installation/), `git`, `curl`. - `ffmpeg`, only for rendering videos.

- The motion data is **not redistributed** here. The GIF, videos and trained checkpoints are derived from LAFAN1. They are provided for non-commercial research and illustration only and are **not covered by this repository's MIT license**. ## Citation If you use InstantTrack, please cite this repository. Please also cite the works it builds on, in particular SuperMimic, InstantMimic, mjlab and LAFAN1: ```bibtex @misc{instanttrack2026, title = {InstantTrack: Whole-Body Motion Tracking for the Unitree G1 in 15 Minutes on One GPU}, author = {Huy-Quang-Dao}, year = {2026},