kornitsovivan/rbdata

AXIS Franka training data for Subnet 80

Per-task LeRobot v3.0 datasets converted from axisrobotics/Franka-Datasets-v2 and axisrobotics/Franka-dataset-v2-2 into the form the Subnet 80 AXIS evaluator feeds a pi0.5 policy. The terms of the source datasets apply to this data.

Layout: <source repo name>/<source task path>/ (one LeRobot v3.0 dataset per task).

	Source	Here
Rate	15 Hz	5 Hz (control rate): frames with frame_index % 3 == 0
Cameras	third_person, wrist, 640x360	both, 224x126 = the content of resize_with_pad(224, 224)
Video	H.264 4:2:0 CRF 30	H.264 High 4:4:4 CRF 16, no B-frames, IDR per episode
state / action	[joint_1..7, finger_1, finger_2]	[finger_1, finger_2, joint_1..7] (evaluator order)
Other columns	ee_state, camera intrinsics/extrinsics	dropped

• action at each kept frame is the source action at that frame (absolute joint-position targets).
• To get the exact policy input: openpi_client.image_tools.resize_with_pad(frame, 224, 224).
• meta/stats.json and per-episode stats are computed on the stored data.
• meta/conversion.json in each task: source revision, parameters, validation (PSNR against the exact model view, frame/timestamp checks) and sha256 of every file.