User Guide · Reference · Task catalogue

Task catalogue

MetaFine registers its envs under simple lowercase ids (grasp_part, peg_in_hole, …). The table below covers the canonical envs that ship with the repo — enough to record demos, train policies, and run the three-dimensional diagnostic eval out of the box.

Canonical envs

Env id	Phase	Default asset	Applicable skills
`grasp_part`	interaction	100221	`grasp_part`
`press`	interaction	103594	`press`
`flip_switch`	interaction	103693	`flip_switch`
`lift_lid`	interaction	100221	`grasp_part`, `lift_lid`
`pure_rotate`	continuation	100221	`grasp_part`, `pure_rotate`
`pure_slide`	continuation	101713	`grasp_part`, `pure_slide`
`pure_insert`	continuation	peg_v1	`grasp_part`, `pure_insert`
`pure_lift`	continuation	cube_v1	`grasp_part`, `pure_lift`
`release_gripper`	continuation	any	`release_gripper`
`slide_along`	bundle	101713	grasp + slide chain
`toggle_switch`	bundle	103693	grasp + flip + release
`peg_in_hole`	bundle	peg_v1	grasp + lift + align + insert
`plug_charger`	bundle	plug_v1	grasp + align + insert
`lid_opening`	bundle	100221	grasp + rotate + lift
`stand_up`	bundle	bottle_v1	grasp + lift + rotate
`draw_triangle`	bundle	pen_v1	grasp + multi-waypoint trace

Phase semantics

interaction — engages an object from rest (grasp_part, press, …).
continuation — operates on an already-engaged part (pure_rotate, pure_slide, release_gripper). Safe to chain after any interaction skill.
bundle — pre-composed multi-step routines kept atomic for record-time stability (lid_opening, stand_up, toggle_switch).

Quick usage

# Record demos for the bundle
$ python record.py --task lid_opening --object 100221 --trials 30

# Or compose your own from atomics via a task graph
$ python record.py --task-graph configs/my_lid_then_rotate.yaml --trials 30

Complete list is auto-generated from gym.registry; this table shows the canonical 12–15 envs that ship with MetaFine. Run python -c "from utils.task_registry import list_envs; print(list_envs())" for the live set.