- Remove fix_dataset.py (user fixes dataset at source)
- evaluate.py: replace observation.pose/joints with observation.state
(8D, derived from action during training, from FK at inference)
- evaluate.py: remove cam1 — training uses only cam0
- docs: rewrite workflow around derive_state_from_action=true,
updated recompute-stats and training commands with
relative_exclude_joints for gripper dims
Made-with: Cursor
- Add replay.py script and replay_viewer.html for browser-based EE
trajectory visualization from glannuzel/grabette-dataset
- Add viewer.html for interactive URDF inspection
- Move OpenArm URDF and meshes into openarm_follower/urdf/
- Add virtual EE target frame (openarm_right_ee_target) at 7cm from link7
- Adapt evaluate.py for single right-arm OpenArm with wrist camera
- Update docs with replay viewer usage
- Update openarm_follower config, driver, and kinematic processor
Made-with: Cursor
Switch from SO100 to a single right OpenArm follower with one camera
(cam0 at 960x720). Strip dataset recording — just execute the policy.
Filter out .vel/.torque observation features for the EE pipeline.
Made-with: Cursor
Pepijn