Towards Online Risk Assessment for Human-Robot Interaction: A Data-Driven Hamilton-Jacobi-Isaacs Reachability Approach

This paper introduces a novel online data-driven methodology for risk assessment in human-robot interaction scenarios. Traditional risk assessment techniques relying on the Hamilton-Jacobi-Isaacs (HJI) method are computationally intensive and require offline calculations. To overcome these limitations, the proposed approach leverages neural networks to predict control actions and construct an augmented dynamic model. The model then utilizes the HJI method to predict the Forward Reachable Set (FRS) for risk assessment. This approach addresses the challenges posed by high-dimensional states by introducing a state-of-the-art neural network as a value approximator. Both the pre-trained motion prediction model and value approximator are aggregated to enable online precise FRS inference. Through comprehensive testing and validation, the effectiveness and feasibility of this methodology are demonstrated in ensuring safety during human-robot interactions. Case studies featuring different motion prediction and dynamic models further validate the practicality of this approach. Furthermore, the feasibility of this data-driven methodology is demonstrated through an online risk assessment of a robot in an intersection scenario.