EdwardoSunny/xarm_cube_points9_arrow_len0
xarm_cube_points9_arrow_len0 xArm7 phone-teleop dataset for cube ("pick up the cube") with the points9_arrow_len0 tactile overlay burned into the camera frames: 9 dots per gripper finger drawn at their real 3D tactile-pad positions (forward kinematics, projected through camera intrinsics), with force arrows rendered at zero length (arrow_length_scale=0). This is the force-ablation control for the points9_arrow variant: identical dot positions and visual salience, but zero force… See the full description on the dataset page: https://huggingface.co/datasets/EdwardoSunny/xarm_cube_points9_arrow_len0.
xarmcubepoints9arrowlen0
xArm7 phone-teleop dataset for cube ("pick up the cube") with the points9_arrow_len0 tactile overlay burned into the camera frames: 9 dots per gripper finger drawn at their real 3D tactile-pad positions (forward kinematics, projected through camera intrinsics), with force arrows rendered at zero length (arrow_length_scale=0).
This is the force-ablation control for the points9_arrow variant: identical dot positions and visual salience, but zero force information in the pixels (verified byte-exact: a zero-force frame and a max-force frame render identically).
Contents
- 100 episodes / 13103 frames @ 10 Hz, LeRobot v2.1 format
image(224x224x3, agent view) andwrist_image(224x224x3) — overlay burned instate(8-dim: eeposm, eeaxisangle_rad, grasp, grasp)actions(7-dim world-frame deltas: dxyzm, daxisangle_rad, grasp in {-1,+1})
Provenance
Raw teleop zarr (RealSense cameras + 2x A31301 3x3 tactile arrays) -> tactile-data-collection/scripts/render_arrowlen0.py (points9arrow mode, arrowlengthscale=0) -> `tactile-openpi/examples/xarm/convertzarrtolerobot.py --mode points9arrowlen0`. Same source recording as every other overlay variant of this task — only the burned-in pixels differ.
Paired model
Trained pi0.5 LoRA checkpoint: EdwardoSunny/pi05_xarm_cube_points9_arrow_len0_lora
