`jointpose` checkpoints quality

[GuyTevet]

Thank you for your great work!

Regarding the jointpos checkpoints, and more specifically the Droid ones:
gs://openpi-assets/checkpoints/pi05_droid/
vs.
gs://openpi-assets/checkpoints/pi05_droid_jointpos/

1. Are they using pose control instead of velocity control?
2. To deploy them, do I need to switch the droid controller type to positions in here:
   

   openpi/examples/droid/main.py

   Line 80 in 981483d

   env = RobotEnv(action_space="joint_velocity", gripper_action_space="position")

3. What config name should I use with it?
4. From your experience, is their quality comparable to the original checkpoint?

[codemypray]

1. if you are using the droid codebase, velocity is converted to delta position by integral, so in fact is pose control though the api is velocity

2. I use a simple convert func to translate pi05_droid_jointpos output into velocity
   def target_qpos_to_droid_qvel_action(
   target_qpos_8d: np.ndarray,
   current_joint_qpos_7d: np.ndarray,
   ) -> np.ndarray:
   """
   Invert DROID's joint velocity -> target qpos mapping approximately.

   Forward mapping in DROID:
   qvel(7) --(global inf-norm scaling to <=1)--> delta_qpos = qvel * 0.2
   target_qpos = current_qpos + delta_qpos

   This inverse computes:
   qvel = (target_qpos - current_qpos) / 0.2
   and applies the same global scaling to keep qvel in [-1, 1].
   """
   target_qpos = np.asarray(target_qpos_8d, dtype=np.float32).reshape(-1)
   if target_qpos.shape[0] != 8:
   raise ValueError(f"target_qpos_8d must have shape (8,), got {target_qpos.shape}")

   current_joint_qpos = np.asarray(current_joint_qpos_7d, dtype=np.float32).reshape(-1)
   if current_joint_qpos.shape[0] != 7:
   raise ValueError(
   f"current_joint_qpos_7d must have shape (7,), got {current_joint_qpos.shape}"
   )

   joint_delta = target_qpos[:7] - current_joint_qpos
   joint_qvel = joint_delta / DROID_MAX_JOINT_DELTA
   max_abs = float(np.max(np.abs(joint_qvel)))
   if max_abs > 1.0:
   joint_qvel /= max_abs

   gripper_position = np.float32(np.clip(target_qpos[7], 0.0, 1.0))
   return np.concatenate([joint_qvel, np.array([gripper_position], dtype=np.float32)], axis=0)

3. pi05_droid_jointpos

4. in my droid setting hardware & software, pi05_droid_jointpos seems to perform way more better than pi05_droid, I am still figure out why...