Learning to predict trajectories of cooperatively navigating agents

Abstract

Abstract-The problem of modeling the navigation behavior of multiple interacting agents arises in different areas including robotics, computer graphics, and behavioral science. In this paper, we present an approach to learn the composite navigation behavior of interacting agents from demonstrations. The decision process that ultimately leads to the observed continuous trajectories of the agents often also comprises discrete decisions, which partition the space of composite trajectories into homotopy classes. Therefore, our method uses a mixture probability distribution that consists of a discrete distribution over the homotopy classes and continuous distributions over the composite trajectories for each homotopy class. Our approach learns the model parameters of this distribution that match, in expectation, the observed behavior in terms of user-defined features. To compute the feature expectations over the highdimensional continuous distributions, we use Hamiltonian Markov chain Monte Carlo sampling. We exploit that the distributions are highly structured due to physical constraints and guide the sampling process to regions of high probability. We apply our approach to learning the behavior of pedestrians and demonstrate that it outperforms state-of-the-art methods.

Predicting actions to act predictably: Cooperative partial motion planning with maximum entropy models

“…In Henry et al [19], the authors extend maximum entropy IRL to work in partially observable dynamic environments and introduce features that capture aspects of crowd navigation. Kuderer, Kretzschmar et al [1,3,4] leverage this approach to continuous state-spaces and introduce features to capture sociallycompliant navigation behavior. They show that the approach outperforms the social forces model [2] and RVO [8] in terms of its predictive qualities of pedestrian motion.…”

“…If the task is to integrate the robot in a smooth fashion into a workspace shared with other agents, optimization of, e.g., travel time will result in aggressive behavior, perceived as unnatural by other individuals. Eventually we wish to realize a robot navigation style which is close (in some sense) to human behavior, commonly referred to as "social compliance" [1]. Otherwise, a navigation behavior perceived as artificial will always stand out and will cause artificial interaction which diminishes the performance of the robot.…”

“…Results from real-world tests are shown in Section V before a conclusion is drawn in Section VI. We highlight deviations from [1,3,4] in the corresponding sections.…”

“…The work presented in this paper builds upon the approach developed in [1,3,4] in which a maximum entropy distribution over joint navigation behavior is learned from demonstration. The contribution of this paper is the real-world application and evaluation of the approach and modifications to the original formulation to minimize the deployment e↵ort in unknown environments.…”

“…The contribution of this paper is the real-world application and evaluation of the approach and modifications to the original formulation to minimize the deployment e↵ort in unknown environments. After the approach showed to outperform state-of-the-art methods in lab environments [1], we determine whether human-like behavior actually improves the navigation performance of the robot among pedestrians unaware of the experiment. We tested the robot's navigation capabilities with both the interaction-aware model and a twostage predict-react motion planner in runs in a public building for several hours.…”