PickCube-AVM: multi-view PickCube with camera poses and state-derived progress

2940 clips of ManiSkill PickCube-v1, each 32 frames at 256x256, rendered from a known camera pose. Built to test whether a progress-reward model conditioned on camera geometry generalises to viewpoints it never trained on.

Clips are stored as individual .npz files in clips/, unarchived and unpartitioned -- pick your own split from the metadata in index.json.

What is in a clip

Each .npz holds frames (T, 256, 256, 3) uint8 and a JSON meta:

field	meaning
progress	per-frame label in [0,1], a pure function of simulator state
cam_params	intrinsic_cv (3x3), extrinsic_cv (3x4), eye, target, fov
states	per-frame tcp, cube, d, lift, src_frame
cube_frac / arm_frac	per-frame fraction of the image covered by cube / robot
kind	success, failure_missed, failure_dropped, recovery

import json, numpy as np
z = np.load("clips/traj_107__success__canonical.npz", allow_pickle=False)
frames, meta = z["frames"], json.loads(str(z["meta"]))

Labels come from physical state, not frame index. Progress is 0.00-0.45 from gripper-to-cube distance, 0.45-0.65 from lift height up to 0.02m, 0.65-1.00 from cube height toward the goal. Replaying the same state later in a clip gives the same label, and a dropped cube's progress falls back on its own.

Geometry is verified, not assumed. Reprojecting the cube's world position through the stored extrinsic_cv and intrinsic_cv lands within ~1 px of the rendered cube. The intrinsic is computed from the FOV actually set, because ManiSkill's sensor_param["intrinsic_cv"] keeps reporting the focal length the camera was registered with and does not follow set_fovy.

index.json

One record per clip -- path, traj, kind, cam, group, mean_cube_frac, progress_range -- plus the full camera battery. Two fields are worth understanding before you split the data:

• group is the camera's role. train means the pose was drawn per trajectory from a continuous distribution (1680 clips, ~140 distinct viewpoints); the other groups are the 21 fixed evaluation poses (1260 clips).
• split records which of the two camera regimes a trajectory was rendered under, not a partition you have to adopt. A trajectory marked train was shot from 12 sampled poses and has no battery clips at all; one marked val/test was shot from the fixed battery and has no sampled ones. So the two regimes are not interchangeable: you cannot ask for a battery view of a train trajectory, because it was never rendered. Regroup trajectories freely within a regime; across regimes, check what exists first.

Cameras

Sampled (per-trajectory) viewpoints: azimuth +-60 deg, elevation 20-55 deg, radius 0.50-0.85 m, FOV 40-60 deg. The fixed battery, held constant across trajectories so that per-view metrics are comparable:

group	cameras	elevation (deg)	|azimuth| (deg)
canonical	1	42-42	0-0
id_random	4	26-45	29-56
ood_pose	8	7-34	86-136
ood_fov	4	24-49	10-58
occlusion	4	12-25	141-158

Trajectories by kind: success 120, recovery 40, failure_missed 20, failure_dropped 20.

Camera vetting

Every battery camera was screened before rendering by replaying trajectories and measuring, from the segmentation, how often the cube is visible. A camera is rejected and resampled if it is blind for more than 25% of frames on its worst probe trajectory, or if the cube covers less than 0.10% of the frame there.

This matters: an earlier version of this battery had 6 of 21 cameras that never saw the cube at all -- two sat inside the robot base, four looked down from above 65 deg elevation, where the arm reaches over the cube and occludes it. Here 1 of 840 battery clips on held-out trajectories fall below 0.1% cube visibility.

Known limitations

• A frame counter is a strong baseline on success. Progress in a successful demo advances monotonically with time, so predicting the mean label per timestep scores Kendall tau +0.97 on success trajectories and +0.86 over the whole labelled set. Only recovery resists it (+0.51). Evaluate there, or report against the frame-counter line explicitly -- a result at or below it is not a result.
• Failures carry no progress target under the usual masking policy, so 40% of the trajectories score only view-consistency metrics.
• Elevation extrapolation is not testable on this task. Above the trained band the arm occludes its own workspace, so the battery varies azimuth only.
• One camera (oodfov01, el 49 / az -58) is blind on a single trajectory; filter on mean_cube_frac if that matters.