Penalize the hands when their palm distance is more than ``distance_margin`` below reset spacing.

Interpolate between initial value iv and final value fv, for any arbitrarily nested structure of lists/tuples in 'data'. Scalars (int/float)

Terminate when NaN/Inf appears in key simulation states. This typically happens when reset randomization spawns interpenetrating bodies (hand/obj

Reward reaching the object using a tanh-kernel on end-effector distance. The reward is close to 1 when the maximum distance between the object an

Reward improvement over the historical closest end-effector distance. Computes the distance between the object and the end-effector bodies (max o

Reward reaching the object using a tanh-kernel on mean end-effector distance. The reward is close to 1 when the average distance between the obje

Terminate if object is farther than `threshold` meters from the Tianji palm.

Object position in the robot's root frame. Args: env: The environment. robot_cfg: Scene entity for the robot (reference frame). D

Object orientation in the robot's root frame. Args: env: The environment. robot_cfg: Scene entity for the robot (reference frame)

Reward cube flight along the initial source-to-receiver y corridor of the bimanual robot.

Reward tracking of commanded orientation using tanh kernel, gated by contact presence.

Termination condition for the object falls out of bound. Args: env: The environment. asset_cfg: The object configuration. Default

Parse configuration for RSL-RL agent based on inputs. Args: task_name: The name of the environment. args_cli: The command line ar

Reward tracking of commanded position using tanh kernel, gated by contact presence.

Persistent print.

Sample or preserve handover commands before reset assets are placed.

Return proprio history length used by ProprioAdapt.

Randomize camera focal length on reset.

Randomize dome-light intensity and color.

Randomize dome-light HDRI texture and yaw rotation.

Randomize camera pose around a fixed base pose.

Reward the receiver-side arm returning to default once 3+ fingertips catch the cube.

Reward the receiver only after at least ``min_fingers`` fingertips reach the cube surface.

Reset all envs and return obs_dict only.

Reset both hands to the default open pose.

Reset object root pose to be `height` meters above the Tianji palm (world +Z).

Reward the throwing-side arm returning to default, with optional light finger regularization after release.

Penalize excessive post-release action-rate on the throwing-side dexterous hand.

Stage A: reward releasing the cube from the source hand, but stop after catch.

Penalize released cubes that stall before reaching the goal region.

Reward success by comparing commanded pose to the object pose using tanh kernels on error.

Orientation tracking reward gated by Tianji hand contact.

Position tracking reward gated by Tianji hand contact.