| --- |
| library_name: pytorch |
| license: other |
| pipeline_tag: robotics |
| tags: |
| - robotics |
| - imu |
| - inertial-odometry |
| - humanoid |
| - human-motion |
| datasets: |
| - bones-studio/seed |
| - Open-IO/humanoid-g1-odometry |
| --- |
| |
| # A Unified Inertial Encoder for Human and Humanoid Motion: Checkpoint |
|
|
| [Project Page](https://unified-inertial-encoder.github.io/) | [Code](https://github.com/unified-inertial-encoder/Unified-Inertial-Encoder) |
|
|
|  |
|
|
| One encoder, trained on human motion and kinematically retargeted G1 motion, transfers zero-shot to |
| a physical Unitree G1. This repository holds the checkpoint behind the paper's "Ours" results. |
|
|
|  |
|
|
| ## Files |
|
|
| ``` |
| checkpoints/unified_encoder.ckpt model weights (EMA, bf16, 55.6M parameters) |
| .hydra/ Hydra config of the training run (experiment=exp15_full_train_C) |
| ``` |
|
|
| Keep `.hydra/` next to `checkpoints/`, because the code rebuilds the model from it. The checkpoint |
| holds weights only, without optimizer state, so it supports evaluation and fine-tuning |
| (`'meta.resume_from=<CKPT>' meta.resume_full_state=false`), not exact resumption of the training run. |
|
|
| ## Quick Usage |
|
|
| With the [code](https://github.com/unified-inertial-encoder/Unified-Inertial-Encoder) installed: |
|
|
| ```bash |
| RUN="${FOUNDATIONIMU_EXP_ROOT:-experiments}/unified-inertial-encoder" |
| hf download Open-IO/Unified-Inertial-Encoder --local-dir "$RUN" |
| hf download Open-IO/humanoid-g1-odometry --repo-type dataset \ |
| --local-dir "$FOUNDATIONIMU_CACHE_ROOT/humanoid_superodom_clips_v2" |
| |
| python -m src.eval_foundation \ |
| --ckpt "$RUN/checkpoints/unified_encoder.ckpt" \ |
| --data_from "$RUN" \ |
| --dataset Humanoid_SuperOdom_NPZ_Processed_Language \ |
| --json real_g1.json |
| ``` |
|
|
| `src.eval_foundation` needs a CUDA GPU. See the code repository for the other datasets. |
|
|
| ## Results |
|
|
| Scores reported in the paper. AVE is the average velocity error (m/s). ATE<sub>GT</sub> and |
| ATE<sub>Pred</sub> are trajectory errors (m) using ground-truth rotation and rotation preintegrated |
| from the gyroscope. |
|
|
| | Dataset | AVE | ATE<sub>GT</sub> | ATE<sub>Pred</sub> | |
| |---|---|---|---| |
| | Nymeria | 0.026 | 2.468 | 18.20 | |
| | Natural Motion | 0.049 | 7.456 | 50.54 | |
| | BONES-Seed (Human) | 0.031 | 0.113 | 0.148 | |
| | BONES-Seed (G1) | 0.024 | 0.101 | 0.127 | |
| | Real-Trajectory (G1) | 0.099 | 3.153 | 11.50 | |
|
|
| Real-G1 clips are never used for gradient training, but their scores informed the selection of this |
| checkpoint. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{chen2026unifiedinertial, |
| title={A Unified Inertial Encoder for Human and Humanoid Motion}, |
| author={Chen, Yutian and Zhao, Shibo and Duan, Rocky and Liu, Karen and |
| Sferrazza, Carmelo and Shi, Guanya and Qiu, Yuheng and Sreenath, Koushil}, |
| year={2026}, |
| url={https://unified-inertial-encoder.github.io/} |
| } |
| ``` |
|
|
