Rewards

Manager: RewardManager · Module: env/mdp/rewards.py

Tracking terms

All primary terms use command_name="motion" and compare robot state to the reference from MotionCommand:

Term	Tracks
motion_global_root_pos / motion_global_root_ori	Root pose vs reference
motion_body_pos / motion_body_ori	Keybody pose (body list from cfg params — set by preset + robot)
motion_body_lin_vel / motion_body_ang_vel	Keybody velocities
motion_joint_pos / motion_joint_vel	Joint state vs reference

Kernel: exp(−κ · error² / σ²). Presets such as apply_wbc set κ, σ, and which bodies participate.

Regularizers

Base template also defines survival, action_rate_l1, joint_acc, joint_limit, actuator_torque_soft_limit, foot_slip, angular_momentum. Presets set weights — many are 0 on default WBC stacks.

Reward-aligned RSI

When RsiCfg.similarity_from_rewards=True, per-step RSI similarity is computed from the same active motion_* terms and weights — see RSI (reference-state initialization).

API: MDP terms (Rewards). Callable names in code use the motion_*_error_exp form (cfg term ids may be shorter aliases in presets).

Related: Tasks and presets, The robot entity.