2019/10/29 by Somil Bansal, Bansal, Somil, Andrea Bajcsy +7 · 2 citations
Computer Science · Engineering · #Gaussian Processes and Bayesian Inference #Anomaly Detection Techniques and Applications #Fault Detection and Control Systems
paper · pdf · doi:10.48550/arxiv.1910.13369
Real-world autonomous systems often employ probabilistic predictive models of\nhuman behavior during planning to reason about their future motion. Since\naccurately modeling human behavior a priori is challenging, such models are\noften parameterized, enabling the robot to adapt predictions based on\nobservations by maintaining a distribution over the model parameters. Although\nthis enables data and priors to improve the human model, observation models are\ndifficult to specify and priors may be incorrect, leading to erroneous state\npredictions that can degrade the safety of the robot motion plan. In this work,\nwe seek to design a predictor which is more robust to misspecified models and\npriors, but can still leverage human behavioral data online to reduce\nconservatism in a safe way. To do this, we cast human motion prediction as a\nHamilton-Jacobi reachability problem in the joint state space of the human and\nthe belief over the model parameters. We construct a new continuous-time\ndynamical system, where the inputs are the observations of human behavior, and\nthe dynamics include how the belief over the model parameters change. The\nresults of this reachability computation enable us to both analyze the effect\nof incorrect priors on future predictions in continuous state and time, as well\nas to make predictions of the human state in the future. We compare our\napproach to the worst-case forward reachable set and a stochastic predictor\nwhich uses Bayesian inference and produces full future state distributions. Our\ncomparisons in simulation and in hardware demonstrate how our framework can\nenable robust planning while not being overly conservative, even when the human\nmodel is inaccurate.\n