# BlinkMimic **Whole-body motion tracking for the Unitree G1 in about 3 minutes on one RTX 4090.** BlinkMimic trains one policy for all 40 LAFAN1 clips (2.45 h of motion) to an evaluation normalized return of 0.82 in **2.9 min of training on one RTX 4090** (mean of 3 seeds), and to 0.82 with 0 of 40 clips failed for 3 checkpoints in a row in **3.4 min**. Like SuperMimic and InstantTrack, BlinkMimic learns one differentiable world model and trains the policy by backpropagation through imagined rollouts, on [mjlab](https://github.com/mujocolab/mjlab) / MuJoCo Warp. BlinkMimic changes how the policy is optimized and where the simulator spends its resets ([details](#what-blinkmimic-changes)). ![Equal training time (3.1 min). SuperMimic and InstantTrack v1.2 fall on fallAndGetUp1_subject1; BlinkMimic tracks it.](docs/static/hero.gif) *Each policy after 3.1 min of training (seed 0), on `fallAndGetUp1_subject1`: the first 8 s, then 30–42 s, simulated in MuJoCo Warp under evaluation conditions. Blue ghost: reference motion. A faded panel has fallen.* ## Results All runs on one RTX 4090 (driver 580, CUDA 12.8, torch 2.9.0), the GPU otherwise idle. 40 LAFAN1 clips; every checkpoint (one per 100 iterations) evaluated on all 40 clips from their first frame with 4 replicas; return and failed clips are averaged over the replicas. Minutes are training wall-clock, including compilation, CUDA-graph capture, probe sweeps and checkpointing, excluding process start-up and data loading. ![Training time to return ≥ 0.82 (left) and to the strict target (right)](docs/static/time_to_target.svg) ### Time to target | | seeds | first return ≥ 0.82 (min) | ≥ 0.82 and 0 failed, 3 checkpoints in a row (min) | iterations to first ≥ 0.82 | ms / iteration | |---|---|---|---|---|---| | SuperMimic | (1) | ≈ 41 (estimated) | ≈ 86 (estimated) | 7898 | 310 | | SuperMimic + EMA | (3) | ≈ 30 (estimated) | ≈ 32 (estimated) | 5749 / 6000 / 5856 | 310 | | InstantTrack v1.2 | 3 | 5.59 / 6.31 / 5.78 (**5.89**) | 7.29 / 8.52 / 6.28 (**7.36**) | 2500 / 2800 / 2500 | 141 | | **BlinkMimic** | 3 | 2.89 / 3.28 / 2.64 (**2.94**) | 3.32 / 3.28 / 3.53 (**3.38**) | 1300 / 1500 / 1200 | 133 | "First return ≥ 0.82" is SuperMimic's metric, the strict column InstantTrack's. ms / iteration = training wall-clock / iterations. SuperMimic times are estimated: the iterations at which InstantTrack's SuperMimic runs reach the target (strict: 16665; 5749 / 6000 / 6971) × 310 ms per iteration, measured for SuperMimic on our RTX 4090. BlinkMimic is 6% cheaper per iteration than InstantTrack v1.2 and needs about half the iterations. With 6 checkpoints in a row instead of 3, BlinkMimic reaches the target in 2 of 3 seeds (5.36, 3.53 min) and InstantTrack v1.2 in 1 of 3 within 4500 iterations. ### Equal training time Mean return / failed clips (of 40) at the last checkpoint within the budget. These SuperMimic runs took 331–348 ms per iteration (310 ms above). | training time | SuperMimic (1 seed) | SuperMimic + EMA (3 seeds) | InstantTrack v1.2 (3 seeds) | **BlinkMimic** (3 seeds) | |---|---|---|---|---| | 3.1 min | 0.179 / 29.2 | 0.002 / 40 | 0.639 / 12.3 | **0.821 / 1.5** | | 4.5 min | 0.367 / 16.8 | 0.231 / 28.9 | 0.767 / 5.0 | **0.851 / 0** | | 7.8 min | 0.495 / 12.2 | 0.480 / 14.0 | 0.846 / 0.1 | **0.861 / 0** | ![Return and failed clips over training time](docs/static/curves.svg) ### Tracking quality at equal training time Mean of 3 seeds, one deterministic rollout per clip (`tools/eval_metrics.py`), at iteration 2000 (4.3–4.6 min) and at the last checkpoint within 7.8 min (v1.2: iteration 3400; BlinkMimic: 3500). | training time | method | return ↑ | failed clips ↓ | body pos (cm) ↓ | body ori (°) ↓ | joints (°) ↓ | foot slip (cm/s) ↓ | contact agreement ↑ | penetration (mm) ↓ | |---|---|---|---|---|---|---|---|---|---| | 4.4 min | InstantTrack v1.2 | 0.765 | 5.3 | 5.15 | 10.3 | 4.85 | 7.75 | 0.746 | 0.079 | | 4.4 min | **BlinkMimic** | **0.851** | **0.3** | **3.74** | **7.37** | **3.99** | **6.87** | **0.757** | **0.061** | | 7.8 min | InstantTrack v1.2 | 0.847 | 0 | 4.07 | 8.29 | 4.77 | 7.18 | **0.775** | 0.061 | | 7.8 min | **BlinkMimic** | **0.862** | 0 | **3.44** | **7.02** | **4.19** | **6.66** | 0.765 | **0.058** | Foot slip of the reference motion itself: 10.2 cm/s. ## What BlinkMimic changes | | SuperMimic ([supermimic_vibed](https://github.com/zita-ch/supermimic_vibed)) | InstantTrack v1.2 | BlinkMimic | flag | |---|---|---|---|---| | policy EMA, resets before recent failure points, contact reward, world model 3 × 1024, learning rate 4e-4 | – | ✓ | ✓ | `PRESET[instanttrack]` | | no host synchronization in the update loop, fused Adam, fused rotations, more of the step compiled | – | ✓ | ✓ | | | Muon on the hidden weight matrices of policy and world model (Jordan et al. 2024; Liu et al. 2025) | – | ✓ | ✓ | `--pi_opt muon --wm_opt muon` | | world-model and policy updates as CUDA graphs | – | ✓ | ✓ | `--cuda_graph` | | goal encoding batched before each imagined rollout (same result, fewer kernels) | – | ✓ | – | | | action built on the reference pose | – | ✓ | ✓ | `--pi_ref_skip` | | policy encoder / decoder | 2048-1024-512-256 | 1024-512-256 | 2048-1024-512-256 | `--enc_hidden --dec_hidden` | | world-model-only warm-up | 500 iterations | 500 iterations | none | `--wm_only_iters 0` | | semi-implicit world model: predicts velocity increments and integrates positions | – | – | ✓ | `--wm_structured` | | left-right mirror augmentation of half of each world-model batch | – | – | ✓ | `--wm_mirror 0.5` | | reset-statistics fixes (duplicate bins; per-clip contact floor of the contact reward) | – | – | ✓ | `--bin_fix --ref_contact_fix` | | evaluation-guided resets: from iteration 1000, every 125 iterations the deployed policy is swept over all clips under evaluation conditions; 15% of the resets go to the 10 s before the places where it fails | – | – | ✓ | `--eprobe_*` | | imagined-rollout batch (policy and critic) | 4096 | 4096 | 2048, policy learning rate 4×, decayed to 1× over 300 iterations once the training return reaches 0.75 (by iteration 1400) | `--pi_batch --pi_lr_mult --lr_decay_*` | | collection step, episode resets and probe sweep as CUDA graphs; sweep on a side stream; checkpoints written by a background thread | – | – | ✓ | `--graph_early --collect_graph --reset_masked --eprobe_graph --eprobe_async --ckpt_light` | The released recipe is `PRESET[main]` in [`scripts/presets.sh`](scripts/presets.sh). `python -m blinkmimic.train` without flags is the SuperMimic baseline; all methods share simulator, MDP, reward and evaluator. ## Setup Linux x86_64, an NVIDIA GPU with a driver for CUDA 12.8, [uv](https://docs.astral.sh/uv/), `git`, `curl`, `python3` (and `ffmpeg` for videos). ```bash git clone https://github.com/fuyichen0725/blinkmimic.git cd blinkmimic uv sync # .venv with mjlab, MuJoCo Warp, Warp, torch 2.9.0 (CUDA 12.8) tools/download_data.sh # 40 G1-retargeted LAFAN1 clips + G1 description -> data/ uv run python tools/test_env.py # sanity check: FK vs simulator, state round-trip, reference playback ``` ## Train, evaluate, reproduce > **Train on an otherwise idle GPU.** On a shared GPU, `--eprobe_async` can crash with an illegal memory access; > add `--eprobe_async 0` there. ```bash bash scripts/train.sh logs/run_s0 --seed 0 # ~8 min: 3500 iterations, a policy checkpoint every 100 uv run python tools/eval_batch.py --run_dir logs/run_s0 --group 18 --reps 4 # 18 checkpoints per batch uv run python tools/time_to_target.py logs/run_s0 bash scripts/reproduce.sh logs/reproduce # the table above: 3 seeds, evaluation, time to target ``` Other tools: `tools/eval_metrics.py` (accuracy and contact metrics), `tools/render.py` (videos from `python -m blinkmimic.evaluate --record`), `tools/export_onnx.py` and `tools/play_onnx.py` (deployment format and a check of it in the simulator; `play_onnx.py` needs `onnxruntime`: `uv run --with onnxruntime python tools/play_onnx.py ...`). ## Attribution and licenses - **Code.** BlinkMimic builds on [supermimic_vibed](https://github.com/zita-ch/supermimic_vibed) (Chong Zhang, ETH Zurich) and [InstantTrack](https://github.com/Huy-Quang-Dao/instantTrack) (Huy-Quang-Dao), both MIT-licensed; their notices are in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md). - **Method.** SuperMimic (Liao et al., arXiv:2610.28480, 2026). InstantMimic (Choi, Leem, Won, arXiv:2609.09821, 2026) for the GPU-native training loop. SuperTrack (Fussell, Bergamin, Holden, 2021) and SHAC (Xu et al., 2022) for learning through a differentiable model. BeyondMimic (Liao et al., arXiv:2508.08241, 2025) for the reward and the evaluation protocol. Muon (Jordan et al., 2024; Liu et al., "Muon is Scalable for LLM Training", 2025). - **Also consulted.** [rethinking-supermimic](https://github.com/btx0424/rethinking-supermimic) (btx0424) and the SuperMimic reimplementations listed in [awesome-supermimic](https://github.com/miao-org/awesome-supermimic). - **Simulation.** [mjlab](https://github.com/mujocolab/mjlab), [MuJoCo](https://mujoco.org), [MuJoCo Warp](https://github.com/google-deepmind/mujoco_warp), [NVIDIA Warp](https://github.com/NVIDIA/warp) (Apache-2.0) and [PyTorch](https://pytorch.org) (BSD-3-Clause). - **Data.** LAFAN1 motions by Ubisoft La Forge (Harvey et al., SIGGRAPH 2020), license **CC BY-NC-ND 4.0**; G1 retargets and robot description by Unitree (BSD-3-Clause), via the Hugging Face dataset `lvhaidong/LAFAN1_Retargeting_Dataset`. The data are not included here. The GIF and trained checkpoints are derived from LAFAN1: non-commercial research use only, not under this repository's MIT License. ## Citation ```bibtex @misc{blinkmimic2026, title = {BlinkMimic: Whole-Body Motion Tracking for the Unitree G1 in Minutes on One GPU}, author = {Chen, Fuyi}, year = {2026}, url = {https://github.com/fuyichen0725/blinkmimic} } ``` Please also cite the works it builds on: ```bibtex @article{liao2026supermimic, title = {SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model}, author = {Liao, Qiayuan and Shao, Yiyang and Long, Junfeng and Zakka, Kevin and Horowitz, Roberto and Abbeel, Pieter and Sreenath, Koushil}, journal = {arXiv preprint arXiv:2610.28480}, year = {2026} } @misc{zhang2026supermimicvibed, title = {supermimic\_vibed: An Unofficial Reimplementation of SuperMimic}, author = {Zhang, Chong}, year = {2026}, url = {https://github.com/zita-ch/supermimic_vibed} } @misc{instanttrack2026, title = {InstantTrack: Whole-Body Motion Tracking for the Unitree G1 in 15 Minutes on One GPU}, author = {Huy-Quang-Dao}, year = {2026}, url = {https://github.com/Huy-Quang-Dao/instantTrack} } @article{choi2026instantmimic, title = {InstantMimic: A High Performance System for Learning Physics-based Skills in Seconds}, author = {Choi, Ikjun and Leem, Geonho and Won, Jungdam}, journal = {arXiv preprint arXiv:2609.09821}, year = {2026} } @article{liao2025beyondmimic, title = {BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion}, author = {Liao, Qiayuan and Truong, Takara E. and Huang, Xiaoyu and Tevet, Guy and Sreenath, Koushil and Liu, C. Karen}, journal = {arXiv preprint arXiv:2508.08241}, year = {2025} } @misc{zakka2026mjlab, title = {mjlab: A Lightweight Framework for GPU-Accelerated Robot Learning}, author = {Zakka, Kevin and Liao, Qiayuan and Yi, Brent and Le Lay, Louis and Sreenath, Koushil and Abbeel, Pieter}, year = {2026}, eprint = {2601.22074}, archivePrefix = {arXiv}, primaryClass = {cs.RO} } @inproceedings{todorov2012mujoco, title = {MuJoCo: A physics engine for model-based control}, author = {Todorov, Emanuel and Erez, Tom and Tassa, Yuval}, booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems}, pages = {5026--5033}, year = {2012} } @misc{mujocowarp2025, title = {MuJoCo Warp: GPU-accelerated MuJoCo with NVIDIA Warp}, author = {{Google DeepMind and NVIDIA}}, year = {2025}, url = {https://github.com/google-deepmind/mujoco_warp} } @article{harvey2020robust, title = {Robust Motion In-betweening}, author = {Harvey, F{\'e}lix G. and Yurick, Mike and Nowrouzezahrai, Derek and Pal, Christopher}, journal = {ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH)}, volume = {39}, number = {4}, pages = {60:1--60:12}, year = {2020} } @misc{jordan2024muon, title = {Muon: An optimizer for hidden layers in neural networks}, author = {Jordan, Keller and Jin, Yuchen and Boza, Vlado and You, Jiacheng and Cesista, Franz and Newhouse, Laker and Bernstein, Jeremy}, year = {2024}, url = {https://kellerjordan.github.io/posts/muon/} } @article{liu2025muon, title = {Muon is Scalable for LLM Training}, author = {Liu, Jingyuan and Su, Jianlin and Yao, Xingcheng and Jiang, Zhejun and Lai, Guokun and Du, Yulun and Qin, Yidao and Xu, Weixin and Lu, Enzhe and Yan, Junjie and others}, journal = {arXiv preprint arXiv:2502.16982}, year = {2025} } ``` ## License MIT ([LICENSE](LICENSE)). Copyright (c) 2026 Fuyi Chen; portions Copyright (c) 2026 Huy-Quang-Dao (InstantTrack) and (c) 2026 ETH Zurich, Chong Zhang (supermimic_vibed). Data-derived files are excluded; see above.