cardboard_box_tcp_curated

Bimanual cardboard-box manipulation demonstrations, collected with a handheld UMI-style fingertip interface and retargeted to end-effector space. This is a curated subset: only whole episodes that passed a per-frame replayability filter (reach, pose, self-collision, joint-velocity) are included. The data is stored as absolute end-effector poses so it can be re-represented by any consumer (Hy-VLA RT-relative rot6d, pi0.5, openpi, or absolute-action policies).

• Task: "Assemble the cardboard box and put it into the bin"
• Episodes: 18 (curated from 38 source episodes)
• Frames: 76,670 at 29.97 fps
• Format: LeRobot v2.1
• Arms: dual (left, right)

Schema

key	shape	dtype	notes
observation.state	(16,)	float32	absolute EE, per arm [x, y, z, qx, qy, qz, qw, gripper]
action	(16,)	float32	same layout; next_state synthesis (the next frame's pose)
observation.images.left_head	(384, 384, 3)	video	left-arm camera
observation.images.right_head	(384, 384, 3)	video	right-arm camera

Per-arm block order is left then right. Quaternions are xyzw (scipy convention) and unit-norm. Gripper is a scalar in [0, 1] (1 = open), kept absolute.

Coordinate conventions (read before using)

• End-effector frame is the tool TCP (the chopstick grasp tip), not the flange.
• Poses live in the episode's ArUco marker/world frame. A -90 deg world yaw about the marker origin is already baked in at conversion - do not re-apply a yaw.
• The camera streams are raw fisheye (Insta360 X5, front_equi), not rectified pinhole. A 384x384 downsample of raw fisheye is out of distribution for standard VLM backbones (PaliGemma/SigLIP). For a plain VLA finetune consider rectifying to a pinhole view first, or feeding fisheye and adapting.

Normalization

stats.json carries min/max/mean/std and quantiles q01/q10/q50/q90/q99 for observation.state and action. The quantiles are over the stored absolute columns, so a policy that quantile-normalizes absolute actions (e.g. pi0.5/openpi on absolute EE) can train directly. Consumers that use a relative action space (Hy-VLA RT-relative rot6d, or a delta-action pi0.5) must compute their own normalization on the transformed values at load time - the stored stats do not describe that space.

Using the Hy-VLA / relative representation

The action is stored absolute on purpose. To reproduce Hy-VLA's representation, apply at load time (not baked into the data):

• state -> 20-d absolute rot6d [xyz(3) + rot6d(6) + gripper(1)] x2
• action chunk -> 20-d RT-relative rot6d [dxyz(3) + relRot6d(6) + gripper(1)] x2, each arm expressed in the anchor pose's wrist frame (T_rel = T_anchor^-1 @ T_i), gripper absolute.

Anchor the action chunk to the current observation.state[t] so train and deploy stay consistent. The RT-relative action is invariant to global world yaw, so the baked -90 deg yaw does not affect action targets - only the absolute state carries the world frame.

Provenance

Curated by a replayability funnel (reach_gate -> data_filter -> export_curation) and re-materialized from source by portable_data_collection.

• Source dataset: byang11259/cardboard_box_tcp @ 42b61515710b4d0f11a8dc4ff1ea532f7cef7413
• Source pdc commit: 9dee003de94221c45f67fbcb3c349a262866610d
• Curation: filter config cardboard_7_17, through stage 3, kept whole episodes only (kept_frac == 1.0); post-process bridge 1.0 s / min length 5.0 s.

Caveats

• One curated task string; not a multi-task dataset.
• Fisheye images are unrectified (see conventions above).
• The gripper has no dynamics/collision model in the source rig, so poses are geometrically feasible but not verified dynamically feasible.