RobotData — procedural-drawer demonstrations

Expert demonstrations for the RoboSynChallenge drawer_open_place task, collected in simulation on procedurally generated drawer assets rather than the single hand-authored drawer the benchmark ships with.

The task is bimanual: the right arm opens the drawer, the left arm picks a tomato and places it inside, the right arm pushes the drawer shut.

Open the drawer, pick the tomato with the other arm then place it in the drawer.

Trajectories are synthesised by an action graph driving IK and TOPPRA motion planning — no teleoperation, no RL. Every episode is one the simulator scored as a success.

Contents

directory	episodes	frames	size	attempts	yield
drawer_ep0_v2/	50	21 250	879 MB	51	98 %
drawer_ep0/	50	21 250	887 MB	50	100 %
drawer_ep0_v2_var4/	50	21 250	872 MB	59	85 %
drawer_v2_variant/	50	21 250	950 MB	345	14 %

200 episodes across four drawers. Yield is episodes saved per trajectory attempted, counted from the collection logs.

All four were collected under identical settings: full domain randomisation (16 events — lighting, camera intrinsics and extrinsics, materials, robot init pose, object poses), two distractor objects drawn from PaperCup / fork / spoon present in every episode, and each drawer at its own true scale.

Format

LeRobot v3.0, robot_type: cobotmagic, 25 fps, 425 frames per episode.

key	shape	contents
observation.state	[14]	joint positions
observation.qvel	[14]	joint velocities
observation.qf	[14]	joint forces — all zero, never written by the simulator
action	[14]	absolute joint targets
observation.images.cam_high	[480, 640, 3]	overhead camera
observation.images.cam_left_wrist	[480, 640, 3]	left wrist camera
observation.images.cam_right_wrist	[480, 640, 3]	right wrist camera

Joint order is [left_arm(6), left_gripper(1), right_arm(6), right_gripper(1)]. Grippers are normalised to [0, 1].

Videos are concatenated per chunk, not per episode — one mp4 per camera holds many episodes. Episode boundaries live in meta/episodes/ (from_timestamp / to_timestamp for video, dataset_from_index / dataset_to_index for the parquet). This is normal for LeRobot v3.0 and differs from v2.1, which writes one file per episode.

Usage

hf download wenjingbian/RobotData --repo-type dataset \
    --include "drawer_ep0_v2/*" --local-dir ./data

ACT and Diffusion Policy read v3.0 directly:

bash policy/act/finetune.sh ./data/drawer_ep0_v2 <output_dir> 0 \
    --steps 80000 --batch-size 64

PI0 / PI0.5 need v2.1:

python scripts/convert_lerobot3.0_to_2.1.py \
    --repo-id drawer_ep0_v2 --root ./data

Training needs nothing but this data. None of the four reference policy adapters reads a URDF, a mesh or a scene config — the model sees images, joint state and joint targets. Simulation assets are published under assets/ for anyone who wants to evaluate a policy in these scenes; see assets/README.md.

The assets

Each directory is named after the procedural asset that produced it.

asset	handle	bar (L × H)	hook gap	tray W	travel	cabinet H
drawer_ep0_v2	protruding bar	130 × 16 mm	3.1 mm	195 mm	180 mm	90 mm
drawer_ep0	protruding bar	130 × 14 mm	2.0 mm	195 mm	245 mm	90 mm
drawer_ep0_v2_var4	protruding bar	169 × 16 mm	3.1 mm	195 mm	170 mm	67 mm
drawer_v2_variant	protruding bar	208 × 20 mm	2.1 mm	312 mm	125 mm	130 mm
benchmark's own drawer	protruding bar	80 × 18 mm	16.5 mm	180 mm	200 mm	200 mm*

* the benchmark models its drawer as a 0.2 m cube and squashes it to 0.07 m via body_scale: [1, 1, 0.35]. Ours are generated at their intended size and are placed unscaled.

Everything downstream of the mesh is derived, not hand-tuned: the grasp point, the collision decomposition, the cabinet's placement, the pull distance and the place target all come from the geometry. The two generators are published under assets/scripts/.

What actually limits yield

Three of the four drawers give a demonstration in 85–100 % of attempts. One, drawer_v2_variant, gives one in 14 % — it needed 345 attempts and 6.8 hours where the others needed 50–59 attempts and about 80 minutes.

drawer_v2_variant is the outlier in size, not in handle shape: a 312 mm tray against 195 mm, a 130 mm cabinet against 67–90 mm, which puts its grasp point at world z = 0.964 against 0.885–0.929 for the rest. Its handle bar is in fact the largest and easiest to grip. The plausible cause is that its geometry pushes the grasp and the pulled-out pose toward the edge of the arm's reachable workspace, but that has not been isolated — treat it as unexplained.

A retraction. An earlier version of this card reported yields of 100 %, 22 %, 22 % and 3.2 % and concluded that handle geometry — bar thickness, hook gap, tray mass — drives demonstration yield across two orders of magnitude. That was wrong. Those numbers came from a bug in our config generator: it inherited the benchmark's body_scale: [1, 1, 0.35] while placing assets that were already the right size, flattening every drawer to a third of its height and dropping the handle 17–42 mm below where the grasp had been solved. Re-collected at true scale, drawer_ep0 went from 3.2 % to 100 %. The handle differences those numbers were attributed to are real, but their effect is small; the squash was doing nearly all of the work.

Anyone using this data should still know that its episodes are survivors of a success filter, not a uniform sample — but the filter is much weaker than that earlier table implied, except for drawer_v2_variant.

Caveats

• observation.qf is declared in the schema but is identically zero. The simulator never writes it, the reference policy adapters never read it, and the benchmark's real-robot datasets do not carry the field at all. Force is not usable through this pipeline.
• 50 episodes per drawer is small for policy training. The benchmark's own guidance is to collect a few hundred per task and merge.
• These are simulation episodes. The benchmark's real-robot data uses a different schema (robot_type: aloha, 32-D state, 10 fps) and needs alignment before co-training.
• A fifth asset, drawer_walnut, is published under assets/ but has no episodes: its handle is a recessed pull whose pocket opening is 70 × 22 mm, and the gripper's fingers separate up to 100 mm. No wrist orientation fits. That is an end-effector mismatch, not a planning failure.