sotac2_weigh_sort_test

Throwaway pipeline test. Not part of the sotac2 dataset family and not for training. Recorded 2026-10-07 to validate the recording stack before the weigh-and-sort data campaign. The real collection lives in Jingyi-Z/sotac2_weigh_sort.

Task

Two visually identical cubes, one light (45.0 g) and one heavy (137.1 g). The robot grasps a cube, holds it under load for 2 to 3 s, then places it directly in the matching box: heavy to the blue box, light to the white box. Left cube first, both cubes sorted in one episode. Task string (identical in every episode):

Pick up each cube and place it in the blue box if it is heavy or the white box if it is light.

Setup

• SO-101 follower arm with two tactile fingertip sensors, teleoperated with an SO-101 leader.
• Four cameras (side_left, side_right, wrist, ego), 640x480, 30 fps, MJPG.
• 30 Hz recording rate.

Data layout

• action: 6 joint targets.
• observation.state: 318 values = 6 joint positions + 2 fingertips x 52 taxels x 3 force components (fixed jaw first). Forces are sensor counts x 0.1 N per count, a vendor nominal with no SI calibration; treat as relative.
• observation.images.*: the four camera streams.
• meta/rig_settings/: camera controls, calibration and software versions captured at recording time.

The session-long high-rate raw tactile sidecar is included under raw/: per-slot magnetics CSVs (~79 Hz unique rate per fingertip), per-slot baselines, and obs_log.csv (epoch-ns timestamps plus a hash of each observation frame, for aligning raw to dataset frames). In the real campaign the raw stream is cut per episode and published in Jingyi-Z/sotac2_raw.