human‑in‑the‑loop RL where a human can take over the robot

Motivation
The Isaac Teleop simulation workflow should support human-in-the-loop reinforcement learning, where a human operator can take over control of the robot mid-rollout. We've been exploring combining Isaac Teleop with RL and have found that human intervention during RL on real robots is relatively straightforward, but reproducing the same kind of intervention in simulation is significantly harder.

The core problem: smooth handover between policy and human control
The difficulty comes down to how cleanly control transfers between the policy and the human at the moment of takeover:

• Real robot (homomorphic teaching arm): The teaching arm's pose can first be matched to the robot's current pose before handing over. Because the poses are aligned at the handover moment, the transition from policy control to human control is smooth.

• Simulation (VR controller): Switching from policy control to human control causes a discontinuity ("jump"). The tracked hand pose typically doesn't match the robot's pose at the moment of handover, so control snaps to the new target. Treating simulation as fast validation before training it on real robot (as it's expensive).

Intended solution: Pause when the policy fails, human takes over from that exact failure state, corrects the trajectory, then resumes. This is what solutions similar to HIL-SERL already do on real robots with a teaching arm.