Kit-Key/libero-3d-scene-pointcloud

LIBERO 3D Whole-Scene Point-Cloud GT (sim, labeled)

Per-frame whole-scene 3D point cloud for the LIBERO benchmark, sampled directly from every object's visual mesh (posed by the replayed MuJoCo state) — complete, view-independent geometry. Every point carries an instance label, and a single is_ooi flag marks the task's objects-of-interest, so relevant objects are extracted with one key at training time.

Scene = task objects + fixtures + the gripper (robot arm/mount excluded). 130 tasks × 50 demos = 6500 episodes, every frame.

Format (one .h5 per episode; per-frame group "%06d")

key	shape / dtype	meaning
scene_pc	f32 (4096, 3)	scene-normalized (scheme B) whole-scene cloud
point_inst	int16 (4096,)	instance index of each point
is_ooi	bool (4096,)	point belongs to an object-of-interest
instance_names	str (K,)	index → instance name
instance_is_ooi	bool (K,)	per-instance OOI flag
object_kinds	str (K,)	graspable | fixture | gripper
ooi_names	str	the task's key objects (from BDDL :obj_of_interest)
norm_params	f32 (4,)	[c(3), r]; metric = pc·r + c
rgb_jpeg	uint8 (N,)	JPEG of the 256×256 agentview RGB

Each OOI instance is guaranteed ≥512 points; the rest of the 4096 budget is split area-weighted across the other instances.

ooi_points = scene_pc[is_ooi]                    # all key-object points (one key)
obj_points = scene_pc[point_inst == i]           # a single instance

Per-object variants: Kit-Key/libero-3d-pointcloud-complete (complete mesh), Kit-Key/libero-3d-pointcloud (single-view depth 2.5D).