Task-format / infra unification (P0–P5): one contract, enforced by guardrail

[geng-haoran]

Tracking issue for the task-format / infrastructure unification effort. Goal: every task on one unified format & infra (BaseTaskEnv/RLTaskEnv, @register_task, scenario: ScenarioCfg, sanctioned _-hooks only, self.handler for sim), deeply integrated and kept aligned by an automated guardrail.

Audit covered 23 families / 362 task files. Canonical contract: MetaSim/metasim/task/{base.py,rl_task.py,registry.py,gym_registration.py}.

P0 — seed contract + guardrail ✅ (done)

• reset(seed) forwarded through 8 family bases (libero, libero_90, maniskill, rlbench, embodiedgen, calvin, pick_place, humanoid) + task_template — fix(tasks): unify reset(seed) contract across task base classes + conformance guardrail #790
• Backend-free conformance guardrail tests/test_task_reset_seed_contract.py (ratchets a 90-file allowlist that may only shrink) — fix(tasks): unify reset(seed) contract across task base classes + conformance guardrail #790
• embodiedgen/put_banana max_episode_steps 250000→250 — fix(tasks): correct embodiedgen put_banana max_episode_steps (250000 -> 250) #791

P1 — eliminate illegal step/reset/close overrides (move logic to sanctioned hooks)

Mechanism: action transforms → pre_physics_step_callback (or an action hook); reward computed in step → _reward; per-reset state → reset_callback/_get_initial_states. As each family lands, remove it from the reset(seed) allowlist and extend the guardrail to ratchet step/close.

step overrides (42): simpler_env (13), pick_place (10), mujoco_playground (3), beyondmimic (3), robosuite (2), mjlab (2), maniskill (2), libero/humanoid_bench/humanoid/dm_control/calvin/benchmark (1 each).
leaf reset-without-seed (90): libero_90 (62 — uniform "skip checker reset" pattern, collapse via a base class flag), pick_place (7), simpler_env/maniskill (4), mujoco_playground/mjlab/beyondmimic (3), robosuite (2), …
close overrides (11): simpler_env (3), robosuite/beyondmimic (2), mjlab/libero/dm_control/benchmark (1).

• libero_90: replace 62 leaf reset overrides with a _skip_checker_reset base flag (biggest mechanical win)
• mujoco_playground (proof-of-pattern): move delta-action scaling → pre_physics_step_callback, reward → _reward, episode bookkeeping → _get_initial_states; drop step/reset overrides
• pick_place: action delta path + tracking-state reset → callbacks; restore canonical reset signature; drop super(PickPlaceBase, self) MRO-skips
• simpler_env / robosuite / maniskill _native: route step through handler.set_dof_targets/simulate/get_states (stop direct scene.step() / mj_step / set_drive_target) — parity-test each (cross-sim obs/reward)
• mjlab / beyondmimic / humanoid / humanoid_bench / calvin / dm_control / benchmark / libero: per-family step/close cleanup

P2 — promote ManagerBasedRVEnv into metasim.task (Tier 2)

The manager-based RL base lives in roboverse_learn.managers; mjlab-v2/beyondmimic/humanoid build parallel manager stacks on it. Promote it to a sanctioned metasim.task base so those families have a blessed lineage; converge mjlab mdp/, beyondmimic mdp/, humanoid CallbacksCfg onto one shared mdp package (dedupe ~60% overlap).

P3 — merge duplicate bases

• LiberoBaseTask ≈ Libero90BaseTask → one base (resolve the inverted checker-reset ordering)
• mjlab v1 vs v2 dual task-IDs → deprecate v1
• beyondmimic metasim vs isaaclab subtrees → one, with a backend selector

P4 — declare the 3 tiers explicitly + unify registration

• Tier 1 native BaseTaskEnv; Tier 2 manager-based; Tier 3 external passthrough (gym registry).
• Give passthrough families (robotwin/metaworld/myosuite/gym_robotics/dm_control/libero_plus/maniskill-v3) thin @register_task adapters so they're discoverable via list_tasks(), and document them as Tier 3.
• Collapse the 3 competing SimplerEnv/ registrars into one; standardize registration prefixes to family.task (fix humanoid unitree_rl./g1./agibot_a2. and beyondmimic's missing prefix).

P5 — config standardization

max_episode_steps as class-var everywhere (not @property/instance/missing); prefer RobotCfg over bare robot strings; migrate DEFAULT_CONFIG-style dicts onto ScenarioCfg/@configclass.

---

Guardrail evolves with each P: the conformance test gains a step/close ratchet and a base-lineage check as the corresponding P lands, so unified format is enforced, not just achieved once.