README.md

metadata

tags:
  - robotics
  - lerobot
  - stringman
task_categories:
  - robotics

simple_grasp_spin

naavox/simple_grasp with the spin state field recovered. Same 370 episodes, same 148,916 frames, same videos - the only difference is one extra number per frame at the end of observation.state.

spin is the gripper camera's heading in the room. The original recording never logged it, which makes the dataset unusable for visual servoing mining: labelling an approach means projecting a room point into the camera, and a camera of unknown heading projects nothing.

How it was recovered

Not estimated - reconstructed, from two things the recording did keep:

1. Every named target carries a bearing, written as room_angle - spin. No target was ever detected in this recording, so they all sit at the room origin, which shows up in the data as distance being exactly the gripper's horizontal distance from it. The room angle is then known from the gripper position alone, so spin = atan2(-x, -y) - bearing is an identity. It degenerates only near the origin, where the direction back to it stops being well defined.
2. spin moves with wrist_angle and nothing else, so the difference between them is a calibration constant. Measuring it on the frames where the bearing is well conditioned fills in the rest exactly.

The constant came out flat to within 0.064 degrees inside every episode, and took three distinct values across the dataset - three calibration sessions.

Coverage

spin is filled for 96.2% of frames, 355 of 370 episodes. The other 15 never left the neighbourhood of the room origin, so their heading could not be measured; their rows carry 0.0 and should be skipped rather than trusted.

Reproducing

python -m nf_robot.ml.visual_servoing.recover_spin \
    --repo_id naavox/simple_grasp --into simple_grasp_spin --videos

from cranebot3-firmware, nf_robot/ml/visual_servoing/recover_spin.py.