Interacting Gaussian Processes

“Unfreezing the Robot: Navigation in Dense, Interacting Crowds”, IROS 2010, Peter Trautman and Andreas Krause.

Abstract: In this paper, we study the safe navigation of a mobile robot through crowds of dynamic agents with uncertain trajectories. Existing algorithms suffer from the “freezing robot” problem: once the environment surpasses a certain level of complexity, the planner decides that all forward paths are unsafe, and the robot freezes in place (or performs unnecessary maneuvers) to avoid collisions. Since a feasible path typically exists, this behavior is suboptimal. Existing approaches have

focused on reducing the predictive uncertainty for individual agents by employing more informed models or heuristically limiting the predictive covariance to prevent this overcautious behavior. In this work, we demonstrate that both the individual prediction and the predictive uncertainty have little to do with the frozen robot problem. Our key insight is that dynamic agents solve the frozen robot problem by engaging in “joint collision avoidance”: They cooperatively make room to create feasible trajectories. We develop IGP, a nonparametric statistical model based on Dependent Output Gaussian Processes that can estimate crowd interaction from data. Our model naturally captures the non-Markov nature of agent trajectories, as well as their goal-driven navigation. We then show how planning in this model can be efficiently implemented using particle based inference. Lastly, we evaluate our model on a dataset of pedestrians entering and leaving a building, first comparing

the model with actual pedestrians, and find that the algorithm either outperforms human pedestrians or performs very similarly to the pedestrians. We also present an experiment where the covariance reduction method results in highly overcautious behavior, while our model performs desirably.

Paper: IROS2010_trautman_krause.pdf

Experimental Results

“Probabilistic Tools for Human-Robot Cooperation”, HART Workshop, HRI, Peter Trautman.

Abstract: We consider the problem of navigating a mobile robot through dense human

crowds.  In particular, we explore two questions.  In the presence of reactive agents, how is a robotic navigation system properly defined?  Can we design a navigation algorithm that encourages humans to treat the robot as an agent with a task to accomplish, instead of as a toy?  In this paper, we address these questions by developing a principled, probabilistic model that predicts how humans perform cooperative collision avoidance, while anticipating their goal-oriented behavior.

Our model extends the recently introduced interactive Gaussian processes approach to the case of multiple goals and stochastic movement duration.  Most importantly, we empirically validate our models in a university cafeteria during peak hours, when the human density reaches 1 person/m^2. Notably, our algorithm performs comparably with human tele-operators, while a simple reactive planner is insufficient for crowd densities above 0.55 people/$m^2$, and a non-interacting planner exhibits freezing robot behavior.  We finish by exploring how this theory naturally extends to a model of human-robot cooperation

Paper: pdf