robocasa-MG_300-pointmap

Depth videos and per-frame camera parameters for the RoboCasa kitchen-24 MimicGen-300 demonstrations, aligned frame-for-frame with DAVIAN-Robotics/robocasa-MG_300-r6. Together the two datasets are the fine-tuning data for the PointmapVLA RoboCasa results. PointmapVLA does not store pointmaps: its data loader back-projects these depth maps with the camera parameters and expresses the result in the robot-base frame at training time.

• 7,200 episodes, 2,075,102 frames, 24 atomic kitchen tasks, 20 fps
• 3 cameras: robot0_agentview_left, robot0_agentview_right, robot0_eye_in_hand
• Rendered from the RoboCasa simulator at 256×256 by replaying the recorded simulator states of the MG_300 demonstrations

Format

LeRobot v3.0. Episodes match robocasa-MG_300-r6 (and the public robocasa-MG_300) one-to-one by episode_index; the parquet files under data/ carry the same action and observation.state columns for convenience plus the camera parameters below. RGB videos are not included here; use the -r6 dataset for RGB, state and actions.

Feature	Shape / type	Meaning
depth.<camera>	video, FFV1 gray16le in .mkv	uint16 = depth[m] × 10000; 0 = invalid. Valid range used by PointmapVLA: 0.05 m < z < 6.5 m. Decode with PyAV in gray16le; RGB video readers destroy the values.
<camera>_camera_intrinsics	float32 (4,)	fx, fy, cx, cy in pixels for the 256×256 render (pinhole, pixel centres at +0.5)
<camera>_camera_extrinsics	float32 (7,)	camera→world pose: position (3) + unit quaternion w, x, y, z (4), OpenCV camera axes, recorded per frame (the wrist camera moves)
observation.state	float32 (18,)	same as -r6: end-effector position (3), rotation as 6-D (first two rows of the rotation matrix) (6), gripper (2), base position (3), base quaternion x, y, z, w (4)
action	float32 (15,)	same as -r6

To obtain a robot-base-frame pointmap: back-project each valid pixel with the intrinsics, transform by the camera→world extrinsics, then by the inverse of the base pose (observation.state[11:14], observation.state[14:18]). The reference implementation is starVLA/dataloader/gr00t_lerobot/transform/pointmap_math.py in the PointmapVLA repository.

Verification

MANIFEST.json lists the size and SHA-256 of every file; check a download with python scripts/verify_data.py <local_dir> from the PointmapVLA repository.

License and attribution

Derived from RoboCasa (MIT License, RoboCasa Team) and its MimicGen-300 demonstrations. The depth renders and camera annotations are released under the MIT License.