emg2tendon — pretrained models

Seq2seq regression from surface EMG to musculoskeletal tendon controls: 16-channel sEMG @ 2 kHz → 39-channel MyoHand tendon control ∈ [0, 1], over 2-second windows (T = 4000). Trained on the full emg2pose dataset (25,253 recordings, 193 subjects, ~370 h), with tendon targets produced by a QForce inverse-dynamics pipeline through the MyoSuite MyoHand model.

Reference implementation + eval code: https://github.com/sagarverma/emg2tendon Project page: https://emg2tendon.github.io Paper: emg2tendon: From sEMG Signals to Tendon Control in Musculoskeletal Hands, Sagar Verma, RSS 2025.

Files

File	Model	Params	val tendon RMSE	open-loop pose (deg)
tds.ckpt	TDS (time-depth-separable conv)	0.10 M	0.310	15.1–16.2
sensingdynamics.ckpt	SensingDynamics	0.56 M	0.308	15.1–16.2
neuropose.ckpt	NeuroPose	7.15 M	0.308	15.1–16.2
cldm.ckpt	Conditional Latent Diffusion (self-contained: both VAEs + U-Net)	7.01 M	0.440	16.8–17.6
emg_stats.npz	per-channel EMG mean/std — required for inference	—	—	—

Checkpoints are PyTorch-Lightning .ckpt files (state_dict + hyper_parameters), loaded by the wrappers in the GitHub repo (RegressionModule for the three baselines, CLDMModule for CLDM).

ablations/ — improvement-campaign checkpoints

Trained on an 8,000-recording subset for speed, to ablate a modernized training recipe (window-sampling fix, velocity + smoothness loss, EMG augmentation, pose-sensitivity-weighted tendon loss, temporal architectures):

File	Description	Params	val tendon RMSE	Δ vs reference
tds_old.ckpt	TDS, original recipe (reference)	0.10 M	0.3105	—
tds_new.ckpt	TDS, new recipe + sensitivity weighting	0.10 M	0.3090	−0.5%
tds_new_nosens.ckpt	TDS, new recipe, no sensitivity weighting	0.10 M	0.3089	−0.5%
tcn_new.ckpt	TCN (~2 s receptive field)	1.57 M	0.3057	−1.5%
gru_new.ckpt	GRU velocity-decode (best)	1.17 M	0.3049	−1.8%

Use ablations/emg_stats.npz with these — the normalization statistics were computed over the 8k subset and differ from the full-25k statistics at the root.

Usage

git clone https://github.com/sagarverma/emg2tendon && cd emg2tendon
pip install torch pytorch-lightning hydra-core diffusers "numpy==1.26.4"

python - <<'PY'
from huggingface_hub import snapshot_download
print(snapshot_download("Micropilot/emg2tendon"))
PY

Then run the shipped eval / render entrypoints:

# pose-space evaluation (per-step + open-loop rollout through MyoHand)
python scripts/evaluate_pose.py --model tds --checkpoint tds.ckpt \
    --index index.json --stats_cache emg_stats.npz

# side-by-side MuJoCo video (reference vs achieved pose)
MUJOCO_GL=egl python scripts/render_model.py --model tds --checkpoint tds.ckpt \
    --emg emg/<base>.npy --pose pose/<base>.npy --index index.json \
    --stats_cache emg_stats.npz --out out.mp4

Inference contract: EMG is per-channel standardized with the shipped mean/std ((emg - mean) / std, guarding std < 1e-8); tendon output stays in native [0, 1] space and is never normalized. Tensors are channel-first ([N, C, T]) at the module boundary — see CONTRACT.md in the code repo.

Evaluation

Protocol: the three held-out emg2pose generalization conditions (unseen user, unseen stage, user+stage), 200 recordings per condition. Predicted tendon controls are forward-simulated through MyoHand and the achieved pose is compared to the emg2pose ground truth, both per-step and in a 0.5 s open-loop rollout.

Model	per-step (deg)	open-loop 0.5 s (deg)
ground-truth tendon (ID ceiling)	0.11	~14.0
TDS / SensingDynamics / NeuroPose	~0.09	15.1–16.2
CLDM	~0.11	16.8–17.6

The inverse-dynamics step itself is near-exact (0.109° MAE per-step over all 25,253 recordings), so per-step numbers sit at the ID ceiling for every model.

Limitations

• The open-loop ceiling is ~14°: even perfect tendon controls drift under MyoHand muscle dynamics. The baselines are already within ~1–2° of it, so there is little headroom for a better EMG→tendon model to improve open-loop pose — these baselines are near-optimal for this metric, not the bottleneck.
• cldm.ckpt is undertrained (diffusion stage cut short for compute budget) and underperforms the regression baselines here — the opposite of the paper's ranking. Treat it as a starting point, not a faithful CLDM result.
• Trained only on emg2pose (wrist-worn 16-ch sEMG band, right/left hand, seated desk-scale gestures). No claims outside that distribution.
• Ablation checkpoints use an 8k subset and are not directly comparable to the root full-25k checkpoints.

License

cc-by-nc-4.0, inherited from the emg2pose dataset these models are trained on (Meta, CC BY-NC 4.0). Non-commercial use only.

Citation

@inproceedings{verma2025emg2tendon,
  title        = {{emg2tendon: From sEMG Signals to Tendon Control in Musculoskeletal Hands}},
  author       = {{Sagar Verma}},
  year         = 2025,
  booktitle    = {Robotics: Science and Systems}
}