Abstract
This paper introduces a fully Bayesian reformulation of Interaction Primitives for human-robot interaction and collaboration. A key insight is that a subset of human-robot interaction is conceptually equivalent to the Simultaneous Localization and Mapping (SLAM) problem. By coupling temporal "phase" estimation and spatial action generation into a single recursive filter, the proposed Bayesian Interaction Primitives (BIP) achieve significant improvements in computational efficiency and accuracy.
Introduction
In human-robot interaction (HRI), a robot must determine both the timing ("when") and the spatial parameters ("how") of its actions. Traditionally, these components have been decoupled. Standard Interaction Primitives (IP) often rely on Dynamic Time Warping (DTW) to align temporal phases before inferring spatial responses. This approach suffers from high computational costs ($O(N^2)$), error accumulation due to decoupling, and poor generalization to varying movement speeds.
Methodology: The SLAM Analogy
The core innovation of BIP is treating HRI as a SLAM problem. In traditional SLAM, a robot estimates its pose while mapping its environment. In BIP:
- Temporal Phase ($\delta$): Acts as the robot's "pose," tracking progress through an interaction.
- Interaction Model ($w$): Acts as the "map," consisting of probabilistic weights describing joint human-robot movements.
The authors implement this using an Extended Kalman Filter (EKF). The state vector includes the phase ($\delta$), phase velocity ($\dot{\delta}$), and basis weights ($w$). This recursive Bayesian approach allows the robot to update its belief about both timing and action simultaneously with each new observation, adapting to mid-action speed changes.
Experimental Evaluation
The BIP framework was evaluated across three scenarios:
- Synthetic 2D Handwriting: Tested robustness against spatial shifts and temporal scaling.
- Human-Robot Handover: A 74-degree-of-freedom (DOF) multimodal task involving a UR5 robot arm, Kinect tracking, IMUs, EMG sensors, and pressure sensors.
- Cooperative Box Lifting: Evaluated system resilience to sensor occlusions where a box blocks the primary visual sensor.
Key Results and Discussion
BIP demonstrated a 66% reduction in computation time compared to DTW-based methods, as EKF updates are significantly more efficient. Furthermore, BIP showed superior accuracy and robustness.
In the handwriting task, BIP was notably more robust to spatial noise. In high-DOF tasks, it maintained lower mean phase error even with partial observations. The framework's Bayesian nature allowed it to handle sensor dropouts gracefully during the box-lifting task by relying on alternative modalities (IMU, EMG) when visual data was occluded.
Conclusion
While BIP offers significant advantages, it remains sensitive to the manual tuning of EKF noise parameters and relies on linear combinations of basis functions. Future work may explore more advanced SLAM techniques, such as Particle Filters or FastSLAM, to handle non-Gaussian scenarios and further enhance interaction complexity.
Summary updated