Probabilistic Dual-Space Fusion for Real-Time Human-Robot Interaction.

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Robots in human environments require sophisticated interaction skills and rapid response to human motion.
• Satisfying both task and joint space constraints in real-time is a significant challenge for robot motion trajectories.
• Hyperspace constraints in human-robot interaction are underexplored, though investigated in robot imitation learning.

Purpose of the Study:

• To enhance the accuracy of inferred robot motion trajectories in both task and joint spaces.
• To develop a method for mapping inferred task space trajectories to joint space trajectories in real-time.
• To create a unified probabilistic framework integrating dual-space fusion, linear mapping, and phase estimation for robot interaction.

Main Methods:

• Proposed a dual-space feature fusion technique to improve trajectory inference accuracy.
• Introduced a linear mapping operator (LMO) to translate task space trajectories into joint space trajectories.
• Developed a unified probabilistic framework combining dual-space fusion, LMO, and phase estimation.

Main Results:

• Dual-space feature fusion significantly improved inference accuracy in both task and joint spaces (>33% over standard Interaction Primitives).
• The second-order LMO demonstrated inference accuracy comparable to kinematic-based mapping methods.
• The unified inference framework achieved a 54.87% reduction in computation time compared to the baseline method.

Conclusions:

• The proposed dual-space feature fusion method enhances robot interaction capabilities by improving trajectory accuracy and real-time performance.
• The unified probabilistic framework offers an efficient and accurate solution for robots operating in human environments.
• This work advances robot learning of complex interaction skills and motion planning under dual-space constraints.