lair-nyu/Pi-Atlas_Sanity_Check_Annotations

Pi-Atlas sanity-check annotations: subtasks and one spatial target per frame

A sidecar to the 640x360 LeRobot v2.1 dataset adityx23/Pi-Atlas_Sanity_Check_Dataset (399 episodes, 270,453 frames, 30 Hz). It holds no images or robot data. Join on (episode_index, frame_index); index (the dataset's global row index) is included and checked.

Each frame gets:

1. A subtask, the instruction at that moment ("pick up the red chili", "go to the workshop", "place the red chili on the plate", ...).
2. One spatial target, the single object that the subtask is about, with its 3D position:

Phase	Subtask	Target
goto	go to ...	the table the robot drives to
pick	pick up the X ...	X (chili or eggplant)
place	place the X on/in the Y	Y: the plate or crate it goes on/in, or the spot on the table where X is put

Files

• frames.parquet: one row per dataset frame, in dataset order
• meta/subtasks.jsonl: subtask_index → text
• meta/segments.jsonl: one record per subtask segment (rows are positions within the episode's map), with the target and the rule that chose it
• meta/episodes.jsonl: per episode, the dataset task, its subtasks and targets, the map it comes from, that map's quality flags and spot-check verdict, and map_clean
• meta/info.json: counts and the code version that produced the release
• subtask_annotations/: the hand annotations as recorded (segment boundaries in frames)

frames.parquet columns

Column	Type	Meaning
episode_index, frame_index, index	int64	join keys, as in the dataset
subtask	string	the instruction ("" for 72 unannotated frames of episode 145)
subtask_index	int64	into meta/subtasks.jsonl, -1 if none
phase	string	goto, pick, place or none
target_label	string	table, chili, eggplant, plate or basket; "" if no target
target_name	string	the map's name for it: plate, table (workshop), table #1, or a spot: table (under chili) = where the chili is put on the table, table (under basket) = the table the crate stands on
target_cam	float32[3]	target in the head (ZED left) camera frame, m: x right, y down, z forward
target_base	float32[3]	target in the robot base frame, m: x forward, y left, z up (approximate: ignores the neck's lever arm)
target_world	float32[3]	target in the map frame (z up, floor at 0). One frame per map, not shared between maps
target_status	int8	0 backfilled, 1 observed, 2 held, 3 unknown, 4 in gripper (see below)
target_valid	bool	there is a target, its status is not unknown, the pose is finite and SLAM tracked
slam_tracked	bool	the camera pose at this frame is a SLAM pose
map	string	the map the row comes from (one per recording; cycle recordings split into laps)
map_clean	bool	the map has no quality flags, or passed the manual spot check

target_status:

• observed: the target was detected near this frame
• held: out of view; its last observed position (or, for a table spot, a point inferred from what stands on it)
• backfilled: before the target's first detection, so its position comes from that first detection. Objects are assumed static until the robot touches them. This uses future information, which a deployed robot doesn't have
• in gripper: grasped and out of view, so it moves with the end effector
• unknown: no estimate; the position is NaN

How it was made

Subtasks come from hand annotation (annotate_subtasks.py) of the three coarse tasks (chili → plate, split into a living-room and a workshop variant; chili → basket, plus 8 reset tails; eggplant → plate). The eggplant cycle episodes are one subtask each already, so they use the dataset task.

Targets come from retroactive object maps, built offline over each whole recording from the full-resolution source dataset:

• ZED depth; YOLO-World detections (SAM 3 text detection for the crate); SAM 3 robot masks
• DROID-SLAM RGB-D over the whole recording, gravity-aligned on the floor, with pose spikes repaired
• detections lifted to 3D and kept when consistent; static objects (tables, plates, crates) pinned to one position; graspable objects attached to the end effector while carried (the neck chain plus recorded end-effector poses)
• the plate or crate the object is put on is identified from where it is released

Maps are then resampled onto these frames by capture time. Each map was checked automatically (tracking, floor fit, grasps, carry error, place error). Flagged maps were then spot-checked by hand.

Caveats

• Positions are estimates. Typical errors: carried objects ~6–9 cm (median, re-sighting after a carry); final place error ~5 cm (eggplant, chili → basket) to ~12 cm (chili → plate). Tables are big: the target is a point on them, not their extent.
• target_world is only meaningful within one map. Use target_cam or target_base for conditioning.
• Backfilled and held statuses are retroactive. Ablate on target_status == observed if test-time targets will only come from live detection.
• map_clean == false rows come from maps with unresolved quality flags; filter on it for the cleanest subset.
• Episode 78 has no chili track, so its 78 pick frames have no target. Episode 145 has 72 frames between annotated segments, with no subtask.

Stats

{
 "dataset": "Pi-Atlas_Sanity_Check_Dataset",
 "created": "2026-11-04T22:41:18",
 "code_commit": "849bf41",
 "frames": 270453,
 "episodes": 399,
 "subtasks": 14,
 "phase_frames": {
  "goto": 176599,
  "pick": 54448,
  "place": 39334,
  "none": 72
 },
 "target_valid_fraction": 0.9938,
 "map_clean_fraction": 0.8971,
 "map_clean_episodes": 361,
 "spot_check": {
  "ok": 11,
  "exclude": 0,
  "fix": 0
 }
}