HIL293 pipette simulation augmentation: 500 demonstrations
500 unique, independently validated simulator demonstrations derived from the
real jren313/g1-pipette-3view-hil293 dataset: 190 rack-placement and 310 smooth
right-arm joint-target augmentations. Robot: Unitree G1 with Inspire hands.
Format and channels
LeRobot v2.1, 30 Hz, one Parquet and three videos per episode.
observation.state: 29 measured simulated motor-joint positions, radians.action: 29 commanded motor-joint targets, radians (not the real 35-D high-level goals).action.hand_left,action.hand_right: separate six-register hand commands.observation.hand_joint_positions: 12 actuated simulated finger angles, radians.observation.images.rgb: fixed ego, 512 x 288.observation.images.wrist_right: fixed side, 384 x 384.observation.images.wrist_left: actual right-wrist camera, 384 x 384. Historical camera key names are preserved; use the roles above. The GROOT modality mapping is inmeta/modality.json.
Augmentation and validation
Rack variants move the rack to put another well underneath the unchanged tip world position. Joint variants add smooth random right-arm target offsets (0.25, 0.5, 1, or 2 degrees), returning to the source targets before insertion. Table height, object sizes, grasp, and camera calibration stay fixed.
All accepted captures pass screening and a separate full, camera-recorded replay: retained attachment, no pipette/ejector-rack contact, tool penetration <= 0.5 mm, normal force <= 100 N, lid/rack penetration <= 0.1 mm, finite physics, no solver warnings, attachment within 0.75 s of the simulated source. No minimum lift. Parent sampling favors sources that yielded successful augmentations; this is not an unbiased evaluation of a learned policy. 1,692 total attempts produced these 500 captures (initial 200 plus 1,492 continuation attempts).
Provenance and training
meta/augmentation.jsonl maps every exported episode to its real source parent,
augmentation recipe, measured lift and validation hashes. Keep all variants and
the corresponding real episode in the same train/validation partition. The
train storage span is not a held-out evaluation split. Earlier HIL293 sessions
also overlap HIL194; account for that overlap when combining datasets.
No remote path is needed to read the Parquet or video files. Absolute paths in provenance describe the collection machine and are informational. Dense replay commands remain in the original capture archive; exported actions are 30 Hz samples at recorded simulation times, not a guarantee of identical 30 Hz replay.
This export is not a drop-in replacement for a 35-D high-level action dataset or
an already-concatenated 41-D motor-plus-hands dataset. Select/concatenate the
separate hand actions explicitly in the training adapter. Do not use privileged
diagnostics.* fields as policy inputs unintentionally.
Limitations
Camera and scene calibration are approximate; the pelvis is fixed and tip attachment uses a geometric constraint. Real-video contact timing is unverified. Some variants derive from episode 102, whose source replay contacts too early. First-session hand commands were reconstructed, and real command timestamps were frozen (commands were paired with image-row times). Simulation success does not establish real-world transfer. No additional dataset license is asserted here.
Ego camera correction — 2026-09-24
The 300 early-session ego videos were rerendered from the saved simulation states using the corrected behind-hand camera. All 500 ego poses now match the authoritative session calibration. Actions, states, outcomes, side videos, and wrist videos are preserved. See meta/ego-camera-repair-audit.json.
Browser video examples
H.264/yuv420p MP4 with fast-start metadata.
- Downloads last month
- 227