Use this checkpoint's own normalization statistics and data config
#1
by akolobov - opened
Fixes https://github.com/microsoft/rhobotics/issues/53 for this checkpoint.
stats.json and data_config.json were copies of rho-base's, made for a TabletopSim dataset. This PR replaces them with this checkpoint's own:
stats.json: statistics of the FR3 Duo dataset, with absolute grippers andq01/q99action-chunk quantiles. Unusedq02/q98entries are removed.data_config.json:repo_id: rho-fr3-duo, this dataset's key names, state and actions read from the quaternion keys and converted to 6-D, anduse_absolute_grippers: true.
The policy's state input and action output now use the 16-value quaternion layout (position, quaternion x/y/z/w and gripper per arm) instead of 20 values with 6-D rotations.
Verified with rhobotics main (d766e66): the checkpoint loads and returns finite actions.
tanreuben changed pull request status to merged