Command terms for the Tianji MDP.

Print per-epoch episode statistics to stdout in addition to TensorBoard logging.

Configuration for randomization.

Wrap Revo3HandHoraEnv for Hora PPO-compatible reset/step signatures.

Observation specifications for the Tianji MDP.

Observations for policy group.

Configuration for recording the data_collection observation group.

Observations for proprioception group.

DirectRLEnv for Revo3 right hand in-hand object rotation. Observation (141 dims) — 3-frame sliding window, 47 dims/frame: [0:21] joint po

Relative joint position control for the public Revo3 arm-hand robot.

Reward configuration dedicated to Revo3 lift/reorient tasks.

Thin async adapter around bc-stark-sdk using radians internally.

Reward terms for the MDP.

Dexsuite Scene for Tianji multi-object lifting.

Build Stage-2 policy inputs from real robot joint positions.

Student proprioception observation group.

Termination terms for the MDP.

Curriculum terms scoped to the Tianji grasp/lift task.

Recorder manager that keeps only action labels and visual-policy inputs.

Command term for bidirectional bimanual handover. The command is a 2D tensor per environment: - command[:, 0]: source-side selector. ``-1`` m

Curriculum terms for the MDP.

Tianji dexsuite lift task definition.

Tianji dexsuite lift evaluation environment definition.

Tianji dexsuite reorientation task base definition.

Tianji dexsuite reorientation evaluation environment definition.

Distillation runner configuration for Revo3 dexsuite tasks.

Configuration for Revo3 lift environment (training).

Configuration for Revo3 lift environment (evaluation/play).

PPO runner configuration for Revo3 dexsuite tasks.

Adaptive difficulty scheduler for curriculum learning. Tracks per-environment difficulty levels and adjusts them based on task performance. Diffi

Uniform pose command generator for an object (in the robot base frame). This command term samples target object poses by: • Drawing (x, y,

Configuration for uniform pose command generator.

Recorder term that saves the ``data_collection`` observation group each step.

Uniform distribution ranges for the pose commands.

Ball variant for Gym registration.

Cylinder variant for Gym registration.

Mixin that attaches the public Revo3 robot asset.

Store episode data in Zarr format compatible with diffusion_policy.

Object surface point cloud expressed in a reference asset's root frame. Points are pre-sampled on the object's surface in its local frame and tra

Reset the kinematic goal marker to the legacy handover target distribution.

Reset the object into a pre-grasp pose relative to the right hand.

Reset the object pose to the IsaacGymEnvs handover start state.