Effect or set_joint_positions with disable dynamics=True.

[mch5048]

Hi, thanks for maintaining this amazing project!

I have question for solving ik via Jacobian that is conducted via rlbench API.

in this line of code we disable dynamics for resetting the agent's position to the initial pose, no matter how the agent is posed at current state.

Meanwhile, if we set the action space to be Carteian space delta, the IK solved action is set to joint position as 'set_joint_target_position' where the dynamics are computed.

I assume this setting yields tremendous errors in doing reinforcement learning with random exploration.

However, in the dm-control environment based on the Mujoco, the mocap based Cartesian space actions seem to work well with the random exploration.

Therefore, I guess we can mimic such behavior of mocap motion in mujoco with set_joint_positions with disable dynamics=True. I want to hear from the author about this remedy.

Thanks!