Jinwei Liu

Shared-control teleoperation systems, which combine human cognitive versatility and robotic precision, are essential for executing complex tasks in inaccessible environments. A critical challenge in such systems is accurately estimating the operator’s intent to ensure the robot’s assistive actions align with the operator’s goals. Traditional trajectory prediction methods estimate the robot’s future path as the operator’s intent. However, without accounting for communication delays, these methods predict the robot’s execution trajectory, which diverges from the operator’s intended trajectory due to the delays. To address this issue, we propose the delay-aware Robot Trajectron (DA-RT), which restores the causal relationship between observed commands and actions by leveraging temporally aligned data and a delay-conditioned attention pooling mechanism. This enables DA-RT to accurately estimate the operator’s intended trajectory under communication delays. Then, we integrate DA-RT with an artificial potential field based controller and a dynamic arbitration mechanism to form a delay-aware shared-control framework, which adjusts the level of assistance based on delay magnitude and the agreement between the operator’s command and the robot’s motion. In human-in-the-loop experiments with fifteen participants, our approach demonstrated significant improvements in both task success rate and efficiency, outperforming traditional methods.

Shared autonomy systems combine user input with autonomous assistance to help users with motor impairments control robot arms for everyday manipulation tasks. However, traditional shared autonomy often moves efficiently but ambiguously toward inferred goals, preventing users from understanding the robot’s internal belief about their intent. We introduce legible shared autonomy, where robot motion simultaneously advances task execution and implicitly communicates which goal the robot has inferred. The method combines an action-level legibility metric with confidence-aware adaptive authority allocation, enabling assertive legible assistance when the robot is confident while preserving user authority under uncertainty. User studies in 2D simulation and physical six degree-of-freedom robot arm experiments demonstrate that legible shared autonomy improves users’ understanding of robot beliefs and reduces user control effort compared with standard shared autonomy.

Neuron-level control offers an effective approach to enabling human-AI shared control without retraining pretrained policies. However, existing methods assume motor primitives are encoded by individual neurons, limiting control precision and robustness. We propose Motor Knowledge Neuron Stimulation (MKNS), which recognizes that motor primitives distribute across neuron assemblies. MKNS identifies these assemblies via frequency-domain analysis and trains compact neural stimulators through iterative data aggregation using only 1% of the original training data. Experiments validate MKNS effectiveness while revealing fundamental properties of motor knowledge: directional asymmetry, the necessity of distributed encoding, and cross-talk between primitives.