Interpretable Goal-based Prediction and Planning for Autonomous Driving

Abstract

We propose an integrated prediction and planning system for autonomous driving which uses rational inverse planning to recognise the goals of other vehicles. Goal recognition informs a Monte Carlo Tree Search (MCTS) algorithm to plan optimal maneuvers for the ego vehicle. Inverse planning and MCTS utilise a shared set of defined maneuvers and macro actions to construct plans which are explainable by means of rationality principles. Evaluation in simulations of urban driving scenarios demonstrate the system's ability to robustly recognise the goals of other vehicles, enabling our vehicle to exploit non-trivial opportunities to significantly reduce driving times. In each scenario, we extract intuitive explanations for the predictions which justify the system's decisions.• A method for goal recognition and multi-modal trajectory prediction via rational inverse planning.

Interpretable Goal Recognition in the Presence of Occluded Factors for Autonomous Vehicles

“…However, these methods lack interpretability which is a challenge for their deployment in trustworthy autonomous systems [16]. The IGP2 algorithm of Albrecht et al [2] is closest to the algorithm we introduce in Section IV; we base inference on inverse-planning for interpretability and then integrate the resulting probabilistic belief into MCTS. However, our work differs in that we infer the presence of potential occluded factors along with goals.…”

“…GOFI outputs a probabilistic belief over possible occluded factor states, Pr z|ŝ i 1:t , and goals of the non-ego, Pr g|z, ŝi 1:t . These beliefs can then be used downstream for ego-vehicle planning, as we show in Section V. GOFI is based on the rational inverse-planning IGP2 algorithm introduced by Albrecht et al [2]. However, since IGP2 does not model occluded factors, it may determine that a vehicle is acting irrationally with respect to one goal even if it is perfectly rational once occluded factors are taken into account.…”

“…At a given moment in time, a controlled vehicle (hereafter called the ego-vehicle) must use a limited number of observations of the current and past states of other (non-ego) vehicles to infer their future locations and velocities, and how they will respond to future ego-vehicle actions. Accurate prediction allows the ego-vehicle to anticipate possible future collisions and plan actions that maximise the probability of safely and efficiently reaching its target location [2], [9], [19], [26].…”

“…Two major approaches to this problem are inverseplanning and deep-learning methods. Inverse-planning methods assume other vehicles behave near-rationally and predict optimal paths under inferred rewards or goals [2], [12]. Deeplearning methods use datasets of past vehicle trajectories and learn the conditional probability of future trajectory segments given past observations and road layouts [10], [14].…”

“…Recently, Albrecht et al [2] showed how goal recognition is important for long-horizon prediction. By inferring the goals (e.g., target locations) that a non-ego vehicle is attempting to achieve and assuming the vehicle behaves approximately rationally, the space of likely future paths for that vehicle is constrained.…”