Teleoperation systems and robot learning.
I build the full stack between a VR operator and a robot — 1 kHz control, low-latency video, the interface, the data pipeline and the policies trained on it — and run user studies to find out what helps the operator and what helps the data.
Now Graduate researcher, MBZUAI · MSc Robotics, Boston University, Dec 2026
Before TU Munich, MIRMI (Haddadin lab) · Tesla · Formula Student
ACT policy trained on my demonstrations, then three rounds of human-gated corrections through the same VR interface. Full sorts 1 → 8 of 20, operator share of rollout time 31 → 14 %. Sim
Policy and DAgger → 70 → 32 NASA-TLX, insertion taskTen-participant study, six assistance conditions, two tasks, ~300 trials on a haptic avatar. Contact force −33 %, no time penalty — and the two assistance channels do not add up. Hardware
Operator study → 415 → 209 ms video latency, p95Every stage of the loop measured on one clock, photon to eye. FEC instead of retransmit, one-frame jitter buffer, velocity feedforward in the controller (−44 ms, predicted −45). Hardware
Latency budget →Policies are only as good as the demonstrations, and through a teleoperation rig people act worse than in person: they cannot feel contact, they see the past, and the interface loads them. Nudging changes what the operator perceives, not what the robot does, so the logged action stays purely human.
Munich showed nudging helps the operator. The current system predicts operator intent live and renders cues from it; the next study trains one policy on nudged demonstrations and one on unassisted ones and compares task success. That study is in design — the human-gated takeover above is its comparator.