Forward unknown attributes to wrapped env.

Initialize the command generator class. Args: cfg: The configuration parameters for the command generator. env: The e

Torque control: torques = p_gain*(target - pos) - d_gain*vel, sent via set_joint_effort_target. p_gain/d_gain are hardcoded from cfg (2.0/0.2)

Termination: height out of [init_z±2cm] or episode timeout (400 steps @20Hz). Gravity curriculum: when reset rate < 0.05%, gravity += 0.05 fro

Compute 6-term reward. Total ×0.01 for PPO. Angular velocity via quaternion difference, no angle gate (pure axis projection).

Delta position control: action ∈ [-1,1] → target += (1/24)*action → clamp to joint limits. Also updates object_rot_prev/pos_prev for angular v

Reset hand to grasp pose (from cache or init_joint_pos), object to default state. PD gains randomized per-DOF each reset: p_gain × [0.5,2.0],

Terminating environment when violation of velocity limits detects, this usually indicates unstable physics caused by very bad, or aggressive actio

Penalize the actions using L2 squared kernel.

Penalize the rate of change of the actions using L2 squared kernel.

Add RSL-RL arguments to the parser. Args: parser: The parser to add the arguments to.

Reward any fingertip contact with the object above threshold. Supports both legacy LEAP sensor names and Arm-BrainCo `right_*` sensor names.

Body state (pos, quat, lin vel, ang vel) in the base asset's root frame. The state for each body is stacked horizontally as ``[position(3), q

Body state (pos, quat, lin vel, ang vel) in the base frame with an optional offset.

Stage C: reward approaching the blue goal point near the receiver.

Stage D: reward a stable low-speed catch near the goal and receiver hand.

The desired pose command. Shape is (num_envs, 7). The first three elements correspond to the position, followed by the quaternion orientation

Disable collisions using USD FilteredPairsAPI based on link groups.

Early-stop episodes once they have been successful for enough consecutive steps.

base-frame contact forces from listed sensors, concatenated per env. Args: env: The environment. contact_sensor_names: Names of c

Stage B: reward forward flight toward the goal with a moderate projected speed band.