CoDance

Data and trained policies for CoDance: Learning Reactive and Compliant Human-Humanoid Interaction from Video (code, project page). The paths mirror the code repository, so downloading into its root makes every tracked config resolve:

uvx --from huggingface_hub hf download GeneralRoboticsLab/CoDance \
    --repo-type dataset --local-dir . --include "data/*"

Clip pools

Path	Clips
data/compliant/rc/20260830_comz_small_002/	400 waltz clips: seeds 0 to 49 are the 200-clip training pool, 50 to 99 the unseen evaluation clips
data/compliant/rc/20260830_stand_comz_small_002/	400 standing clips: seeds 0 to 99 are the training pool

Each seed is one augmented clip (*_multilink_contact_seed<k>.npz plus *_multilink_forcefield_seed<k>.npz) and one zero-wrench clip (*_zerowrench_contact_seed<k>.npz). The rc.yaml in each directory records the source clips, seeds and generator flags.

Evaluated policies

data/checkpoints/<policy>.pt, one per policy in the paper's tables and figures. Each file holds the policy's actor, critic (and discriminator where the policy uses the adversarial style term), the optimizer state, and the frozen train and eval config trees the run was launched with. The training config of the same name is a whole file in the repository under src/mjlab/tasks/codancing/config/g1/conf/<policy>.yaml.

Next to each checkpoint, data/checkpoints/<policy>.onnx is its actor exported for deployment, with the observation normalization baked in and the observation layout and control constants in the ONNX metadata. The real-robot runners in src/mjlab/tasks/codancing/simple_deploy/ load these files.

Checkpoint	Policy
decoupled.pt	the decoupled policy
decoupled_wo_adversarial.pt	decoupled, without the adversarial style term
decoupled_wo_chaining.pt	decoupled, trained without clip chaining
wholebody.pt	the whole-body compliant policy
wholebody_wo_foot_constraint.pt	whole-body, without the foot constraint
stiff.pt	the stiff tracker
stand.pt	the standing policy