Self-contact cost

Use this cost to discourage unwanted contact between robot parts. Choose which bodies to penalize and which other bodies they should not touch.

The BM tracker in examples/experiments/mimic/mlp_bm_l2c2.py scores non-foot/non-hand bodies against all robot links. For G1:

from protomotions.envs.mdp_component import MdpComponent
from protomotions.envs.rewards.meta import RewardMeta
from protomotions.envs.rewards.regularization import compute_self_contacts

allowed = {
    "left_ankle_roll_link", "right_ankle_roll_link",
    "left_wrist_yaw_link", "right_wrist_yaw_link",
    "left_rubber_hand", "right_rubber_hand",  # Noncolliding hand markers.
}
names = robot_cfg.kinematic_info.body_names
bodies = [name for name in names if name not in allowed]
partners = [f"robot/{name}" for name in names]

# In configure_robot_and_simulator():
robot_cfg.update_fields(
    contact_bodies=bodies, contact_partners=partners,
)

# In env_config(), after constructing cfg:
cfg.reward_components["self_contact"] = MdpComponent.from_kernel(
    compute_self_contacts,
    meta=RewardMeta(weight=-0.1),
    # The build refuses a selection the simulator does not measure.
    required_keys={"contact_body_index": bodies, "contact_partner_index": partners},
    body_names=bodies,
    partner_names=partners,
    force_threshold=1.0,  # Normal force in newtons.
)

The simulator measures these contacts; the reward selects which ones to use. A missing requested measurement raises an error at setup. Ground and object contacts do not contribute. Each selected body adds at most one penalty per step, even if several other links touch it. A hand touching the torso penalizes the torso; two knees touching penalizes both knees. Hands and feet are exempt from scoring, not removed as partners. G1 measures 27 rows and 33 columns.

This follows the scored-body selection in BeyondMimic and Holosoma. SONIC additionally exempts elbows; this BM example does not. Those upstream terms use net-contact history, whereas this cost intentionally uses only current robot-to-robot pairs, not terrain or objects.

Train the example with:

python protomotions/train_agent.py --robot-name g1 --simulator isaaclab \
    --experiment-path examples/experiments/mimic/mlp_bm_l2c2.py \
    --motion-file data/motion_for_trackers/g1_bones_seed_mini.pt \
    --num-envs 4096 --batch-size 4096 --experiment-name g1_bm_self_contact

The robot asset must allow self-collisions; the reward does not enable them. Links without collision geometry cannot trigger the cost, and PhysX filters contacts between directly connected parent/child links. For simulator support, see Partner-Resolved Contacts.