Learning Multi-Agent Navigation from Human Crowd Data

• Main Author: Joshi, Foram
• Format: Text
• Language: unknown
• Published: Clemson University Libraries 2021
• Subjects: Orca
• Online Access: https://tigerprints.clemson.edu/all_theses/3555; https://tigerprints.clemson.edu/cgi/viewcontent.cgi?article=4562&context=all_theses

The task of safely steering agents amidst static and dynamic obstacles has many applications in robotics, graphics, and traffic engineering. While decentralized solutions are essential for scalability and robustness, achieving globally efficient motions for the entire system of agents is equally important. In a traditional decentralized setting, each agent relies on an underlying local planning algorithm that takes as input a preferred velocity and the current state of the agent's neighborhood and then computes a new velocity for the next time-step that is collision-free and as close as possible to the preferred one. Typically, each agent promotes a goal-oriented preferred velocity, which can result in myopic behaviors as actions that are locally optimal for one agent is not necessarily optimal for the global system of agents. In this thesis, we explore a human-inspired approach for efficient multi-agent navigation that allows each agent to intelligently adapt its preferred velocity based on feedback from the environment. Using supervised learning, we investigate different egocentric representations of the local conditions that the agents face and train various deep neural network architectures on extensive collections of human trajectory datasets to learn corresponding life-like velocities. During simulation, we use the learned velocities as high-level, preferred velocities signals passed as input to the underlying local planning algorithm of the agents. We evaluate our proposed framework using two state-of-the-art local methods, the ORCA method, and the PowerLaw method. Qualitative and quantitative results on a range of scenarios show that adapting the preferred velocity results in more time- and energy-efficient navigation policies, allowing agents to reach their destinations faster as compared to agents simulated with vanilla ORCA and PowerLaw.