kiteml/trossen-bin-picking-sim-baseline

Trossen bin picking, simulated: baseline

The reference look, lit to match the recordings. One of seven datasets that replay the same recordings under different conditions; see trossen-bin-picking-sim-datasets for all of them, a preview and the limits.

Each episode replays one recording from Trossen Robotics' public bin_picking dataset on a simulated copy of its station. A bimanual WidowX AI picks up one of five objects (red cup, yellow duck, purple marker, screwdriver, roll of black tape) and puts it in a clear tray. The joint states, actions, timing and task text are the recording's, unchanged. The four camera videos are rendered.

Upload in progress: 51 of 94 episodes so far. More are added as they finish rendering.

51 episodes: 46 for training, then 5 held out for testing (splits in meta/info.json). meta/provenance.json gives each episode's source episode, object and outcome.

Key	Shape	Content
observation.images.camera_high	480 x 640 x 3, 30 fps, AV1	Overhead camera
observation.images.camera_low	480 x 640 x 3, 30 fps, AV1	Low front camera
observation.images.camera_left_wrist, observation.images.camera_right_wrist	480 x 640 x 3, 30 fps, AV1	Wrist cameras
observation.state	14	Joint positions from the recording, left arm then right
action	14	Joint targets from the recording

LeRobot v3.0, the same format, keys and robot type (bi_widowxai_follower_robot) as the source.

Source and license

Derived from Trossen Robotics' public bin_picking dataset (s3://trossen-public-datasets-977e8d43bcff/bin_picking/), released under Apache-2.0. This dataset keeps that license. Made by Kite.