Enhancing Teleoperation with Structural Causal Models
Research explores multimodal teleoperation and its implications for robotics.
Recent research by Jinting Hang and Zhenhui Cai investigates the complexities of teleoperated demonstrations in robotics, highlighting the influence of operator habits, shared physics, and observation nuisances on action selection.
The study introduces a structural causal model to analyze these factors and tests its effectiveness through various interventions. Results indicate that modifying actions at a fixed state significantly increases next-state errors, while changes in appearance and camera settings do not have the same effect.
Additionally, the research proposes a habit-aware reverse scoring method that enhances the ranking of feasible past actions without altering the underlying dynamics. This adaptation rule has shown to improve low-shot transfer performance across different robotic platforms, including StackCube, DROID, and RH20T.