Question about rotation representation choice in franka_env.py

[Triple-ROCK]

In the latest issuse fixed, franka_env.py's step calculates nextpos using rotation vector instead of Euler angles, which said euler angles as delta action failed to converge. But why representing state using eulers instead of quat? I believe quaternions would actually be suitable for state representation (as network input), since:

• The double cover issue (q and -q representing the same rotation) only affects network outputs, not inputs
• As observations, quaternions are already normalized and singularity-free
• The network can naturally learn the symmetry between q and -q

Additionally, delta action is supposed to be small increments, which is fine to use euler angles I think, because it will never reach singularity in 1/10 second.

[liusong-0086]

In the latest issuse fixed, franka_env.py's step calculates nextpos using rotation vector instead of Euler angles, which said euler angles as delta action failed to converge. But why representing state using eulers instead of quat? I believe quaternions would actually be suitable for state representation (as network input), since:

• The double cover issue (q and -q representing the same rotation) only affects network outputs, not inputs
• As observations, quaternions are already normalized and singularity-free
• The network can naturally learn the symmetry between q and -q

Additionally, delta action is supposed to be small increments, which is fine to use euler angles I think, because it will never reach singularity in 1/10 second.

1. For SERL’s demo scenarios, using quaternions, Euler angles, or rotation vectors makes almost no difference, because there is only a single degree of freedom—Yaw. A similar issue exists here as well: using linear interpolation for rotations is absolutely incorrect in robotics. However, for a task with only planar degrees of freedom, it can still work.

# /serl_robot_infra/franka_env/envs/franka_env.py
def interpolate_move(self, goal: np.ndarray, timeout: float):
      """Move the robot to the goal position with linear interpolation."""
      if goal.shape == (6,):
          goal = np.concatenate([goal[:3], euler_2_quat(goal[3:])])
      steps = int(timeout * self.hz)
      self._update_currpos()
      path = np.linspace(self.currpos, goal, steps)
      for p in path:
          self._send_pos_command(p)
          time.sleep(1 / self.hz)
      self.nextpos = p
      self._update_currpos()

2. Euler angles are replaced by rotation vectors (Lie algebra) , because Euler angles cannot be multiplied by the adjoint matrix.
3. If your initial pose is close to a singularity, it’s also possible to reach the singularity within 0.1 seconds.