Task

Task (core/task.py) is the abstract base bundling everything task-specific: the DRManager, the robot descriptor, sensor configs, and the Gym action and observation spaces.

Class attributes

Attribute	Purpose
uid, label, description	Identity and human-readable name
metadata	physics_dt (0.002), render_hz, split
robot_cfg	Robot uid and controller selection
sensor_cfgs	Camera definitions (eye_on_base, eye_in_hand, head)
dr_cfgs	Per-randomizer configuration, see Domain Randomization

Key methods

• reset(seed, options) — sequences through every registered randomizer (scene, target, container, distractor, spatial, lighting, camera, material) and builds a Layout that both simulators ingest to rebuild the scene.

• check_success() / compute_reward() — task completion and reward.

• instruction — the natural-language instruction for this episode, sampled by LanguageDR.

• decompose() — splits the objective into SubtaskSpec atoms, used by the motion-planning data generator.

• state_dict() / load_state_dict() — serialize an episode’s layout so it can be reproduced exactly during replay or evaluation.

Concrete tasks live in simple/tasks/ and are named <robot>_<skill>_<pipeline>.py, where the pipeline suffix is _mp (motion planning) or _teleop (VR teleoperation).