aravindsuresh16/smolvla_cathlab_fr3_instruments_v2_5
Robotics • 0.5B • Updated • 23 • 1
640 episodes · 30 fps · 3 cameras · 224×224 av1
This dataset was created using LeRobot.
Dataset Description
v2.5 of the cath-lab instrument dataset: the same 640 episodes as aravindsuresh16/cathlab_fr3_instruments_v1, with the acting arm named in every instruction, e.g. "Hand me the trocar with the left arm" (18 instructions instead of 9). Images, states and actions are identical; only the task strings and task_index differ. The recorded acting arm matched the arm that moves in the actions in 640/640 episodes.
Why: a SmolVLA model trained on the original instructions chose the arm at random (22/48 correct on held-out scenes), because the arm is never stated and depends on which side of the table the named instrument lies. Trained on this version, the model picked the correct arm in 48/48 episodes. Results: docs/V2_RESULTS.md.
Synthetic demonstrations of a bimanual Franka FR3 Duo handing over surgical instruments in a simulated cath lab: "Hand me the {scissors | trocar | puncture device} with the {left | right} arm". The arm on the instrument's side picks it up and places it on a central instrument tray. Physics and rendering: NVIDIA Newton 1.6 (MuJoCo-Warp backend); a scripted IK expert planned on CPU MuJoCo from ground-truth poses. Only successful episodes are kept.
640 episodes, 30 fps, 3 cameras (224 × 224):
| Start | Episodes | What it shows |
|---|---|---|
| Home | 270 | Both arms at the home pose |
| Near target | 270 | The acting arm starts 6–18 cm above and up to 8 cm around the instrument (recovery from imprecise approaches) |
| Drop recovery | 100 | The instrument was dropped mid-lift or mid-carry and lies where it fell; the arm starts raised with the gripper open, then re-picks and places it. The drop itself is not recorded |
Variation: 45 training layouts (4 instruments, two per side, random position and full rotation; tweezers are a distractor), 4 table colours, light intensity 0.7–1.3× and colour temperature 4000–7500 K, object mass ±30 %, friction 0.5–1.0, 3 instruction phrasings × 2 arms. Held out for evaluation (not in this dataset): 5 layouts and 1 table colour. Fixed: tray position and place spot per arm, light direction, camera poses.
Schema: observation.state and action are float32[16] absolute joint targets (left arm joints 1–7, left gripper, right arm joints 1–7, right gripper; gripper 1 = open), matching the FR3 Duo 16-D layout. Cameras: external_camera, left_wrist_camera, right_wrist_camera. meta/episodes_extra.jsonl holds the full configuration of every attempt (layout, start type, acting arm, drop details); manifest.yaml summarises the dataset.
Code: https://github.com/aravinds1610/mujoco_cathlab_vla
Asset credit and licence: the surgical instrument and tray models are from the NVIDIA Isaac for Healthcare asset catalog (https://github.com/isaac-for-healthcare/i4h-asset-catalog, Lightwheel assets), licensed CC BY-NC 4.0. This dataset renders those assets and is therefore released under CC BY-NC 4.0: non-commercial use only, with attribution. Robot models: MuJoCo Menagerie (Franka FR3 v2, Robotiq 2F-85).
- Homepage: [More Information Needed]
- Paper: [More Information Needed]
- License: cc-by-nc-4.0
Dataset Structure
{
"codebase_version": "v3.0",
"fps": 30,
"features": {
"observation.state": {
"dtype": "float32",
"shape": [
16
],
"names": [
"left_joint1",
"left_joint2",
"left_joint3",
"left_joint4",
"left_joint5",
"left_joint6",
"left_joint7",
"left_gripper",
"right_joint1",
"right_joint2",
"right_joint3",
"right_joint4",
"right_joint5",
"right_joint6",
"right_joint7",
"right_gripper"
]
},
"action": {
"dtype": "float32",
"shape": [
16
],
"names": [
"left_joint1",
"left_joint2",
"left_joint3",
"left_joint4",
"left_joint5",
"left_joint6",
"left_joint7",
"left_gripper",
"right_joint1",
"right_joint2",
"right_joint3",
"right_joint4",
"right_joint5",
"right_joint6",
"right_joint7",
"right_gripper"
]
},
"observation.images.external_camera": {
"dtype": "video",
"shape": [
224,
224,
3
],
"names": [
"height",
"width",
"channel"
],
"info": {
"video.height": 224,
"video.width": 224,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"has_audio": false,
"video.g": 2,
"video.crf": 30,
"video.preset": 12,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {},
"is_depth_map": false
}
},
"observation.images.left_wrist_camera": {
"dtype": "video",
"shape": [
224,
224,
3
],
"names": [
"height",
"width",
"channel"
],
"info": {
"video.height": 224,
"video.width": 224,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"has_audio": false,
"video.g": 2,
"video.crf": 30,
"video.preset": 12,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {},
"is_depth_map": false
}
},
"observation.images.right_wrist_camera": {
"dtype": "video",
"shape": [
224,
224,
3
],
"names": [
"height",
"width",
"channel"
],
"info": {
"video.height": 224,
"video.width": 224,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"has_audio": false,
"video.g": 2,
"video.crf": 30,
"video.preset": 12,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {},
"is_depth_map": false
}
},
"success": {
"dtype": "bool",
"shape": [
1
],
"names": null
},
"timestamp": {
"dtype": "float32",
"shape": [
1
],
"names": null
},
"frame_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"episode_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"task_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
}
},
"total_episodes": 640,
"total_frames": 230558,
"total_tasks": 18,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
"robot_type": "fr3_duo",
"splits": {
"train": "0:640"
}
}
Citation
BibTeX:
[More Information Needed]
- Downloads last month
- 113
Models trained or fine-tuned on aravindsuresh16/cathlab_fr3_instruments_v2_5