README.md

metadata

license: apache-2.0
viewer: false

Synthetic bimanual manipulation demonstrations

Bimanual demonstrations collected in simulation on procedurally generated assets: four drawers and two test-tube racks, 1 000 episodes each.

• Drawer: the right arm opens a drawer, the left arm picks up a tomato and places it inside, the right arm pushes the drawer shut.
• Tube into rack: the right arm picks up a test tube and hands it to the left arm, which inserts it into a well of the rack.

Trajectories are synthesised by an action graph driving IK and time-optimal motion planning — no teleoperation, no reinforcement learning. Every episode is one the simulator scored as a success.

Contents

directory	episodes	frames	size	attempts	yield
drawer_ep0_v2/	1 000	425 000	17.9 GB	1 024	97.5 %
drawer_ep0/	1 000	425 000	17.8 GB	1 075	92.6 %
drawer_ep0_v2_var4/	1 000	425 000	17.4 GB	1 374	72.6 %
drawer_v2_variant/	1 000	425 000	18.7 GB	1 124	88.9 %
sample_loading_twin/	1 000	372 000	11.3 GB	1 095	91.3 %
sample_loading_var1/	1 000	372 000	10.5 GB	1 148	87.1 %

6 000 episodes, 2.44 M frames, ~94 GB. Yield is episodes saved per trajectory attempted, from the collection logs.

All six use the same settings: full domain randomisation (16 events — lighting, camera intrinsics and extrinsics, materials, robot initial pose, object poses), distractor objects present in every episode, each asset at its own true scale, 640×480 cameras.

Format

LeRobot v3.0, 25 fps, 14-DoF bimanual arm. 425 frames per drawer episode, 372 per rack 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
duck_pose, drawer_pose	[4, 4]	drawer sets: object poses; drawer_pose is all zero
rack_pose, cube_pose	[4, 4]	rack sets: rack and tube poses

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.

Usage

hf download <this-repo> --repo-type dataset \
    --include "drawer_ep0_v2/*" --local-dir ./data

ACT and Diffusion Policy read v3.0 directly; PI0 / PI0.5 need a v2.1 conversion first. Training needs nothing but this data — the reference policy adapters read images, joint state and joint targets, never a URDF or a scene config. Simulation assets are published under assets/ for anyone who wants to evaluate a policy in these scenes.

Merging these files needs a patch. Stock LeRobot dataset merging fails on them: the 4×4 pose features come back from HuggingFace datasets as an extension arrays that pd.concat cannot hash. Sanitise such columns to nested lists at every read / write / concat boundary first.

The assets

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

"Bar H" includes the end caps; the bar itself is 7–9 mm thick on all four.

Everything downstream of the mesh is derived rather than hand-tuned: the grasp point, the collision decomposition, the cabinet's placement, the pull distance, the place target and the release point all come from the geometry. The two generators are published under assets/scripts/.

Three quantities were originally fixed constants copied from a reference drawer, and each had to become a derived rule before the generated assets worked:

• Pull distance. A single hard-coded value for every drawer, regardless of its rail travel — any drawer with less usable travel jammed against its stop.
• Grasp point. "2 mm behind the handle bar" assumes room behind the bar for the fingertips. One drawer has 2.1 mm of gap; its fingertip hit the door panel, the close stalled, and the bar slipped out on the pull. The rule now keeps 4 mm of clearance from the panel, a property of the gripper rather than the drawer.
• Release point. The gripper opens at a node that carried literal offsets x = 0.18, z = 0.14. Against one drawer's tray that sat 25 mm from the front lip — the tomato is 50 mm across — and 17 mm above the rim, so the tomato settled in or toppled out depending on the pose randomisation. Anchoring it in the closed frame (|pull| + 40 mm, tray centre, rim + 74 mm) took that asset's yield from 55 % to 89 %.

A fourth trap was not a constant but a decomposition setting: convex decomposition thresholds are relative to the whole input mesh, so a 7 mm handle bar on a 200 mm tray fell below the threshold and came back merged with its end caps as a single hull up to 34 mm thick. The gripper closed on that hull and the finger pads sat ~4 mm clear of the visible bar in every wrist-camera frame. Decomposing the handle separately fixes it.

The racks

asset	size (L × W × H)	wells	well opening	tube clearance	plates
sample_loading_twin	275 × 116 × 79 mm	5 × 12	18.5 mm	1.7 mm	3
sample_loading_var1	232 × 128 × 95 mm	4 × 7	20 mm	3.2 mm	3

The twin was generated from a frame of a real rack; the variant is the same parametric model with its dimensions and well grid changed. The tube is 16.8 mm across and 195 mm long. Published under assets/rack_twin/ and assets/rack_var1/.

Three things had to be right before the insertion worked:

• Three plates, not two. A tube only stands upright when it passes through two guide plates; the lean it can take is clearance / plate spacing. With a single well plate it leaned 20–45° and failed the 10° fall check every time.
• Aim at a well on both axes. An even well count leaves a web, not a well, on the centreline. The insertion offset is the nearest well centre along the length and across the width.
• Aim the tube, not the gripper. After the hand-over the tube sits 1–5° off the gripper's nominal axis, differently every episode, so its tip — 97 mm below its centre — is up to 6 mm off. Successes had 1.2° and 3.2° in hand, failures 3.5° and 4.6°. Once the hand-over is finished the tube's actual pose in the gripper is read and the remaining left-arm waypoints are re-solved so the tube is vertical over the well. Yield went from about 45 % to 91 % on the twin and from about 73 % to 87 % on the variant.

What limits yield

drawer_ep0_v2_var4 is the weakest at 72.6 %, and the cause is known but unfixed: over-travel into the hard stop, not push-back timing. Tracing the push shows it finishing with 31 of its 50 steps to spare, while peak joint travel is bimodal — either ~0.130 or exactly 0.1697, its limit — and after hitting that stop the drawer rebounds. Reducing the pull does not help: at pull fractions 0.80 / 0.70 / 0.60 the yields are 83 % / 80 % / 15 %, since a short pull leaves the tomato outside the success radius once the drawer is closed.

Episodes here are survivors of a success filter, not a uniform sample.

Caveats

• observation.qf (all sets) and drawer_pose (drawer sets) are declared in the schema but are identically zero. The simulator never writes them.
• These are simulation episodes only.
• 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.