PI: Ding Zhao
University: Carnegie Mellon University
Companies such as Amazon and Nuro are increasingly deploying autonomous delivery robots across the country to transport food and life necessities during the COVID-19 pandemic. Pennsylvania is among the earliest states to legalize these robots as personal delivery devices (PDDs). As the number of PDDs increases, encounters between pedestrians and PDD will increase, which may hurt the delivery efficiency reduction and cause safety issues. The low-cost requirement for a single robot’s sensing and computing hardware requires smarter methods to collaboratively use information in a network of PDD fleets. In this proposal, our team aims to develop a robust sensing mechanism for a network of autonomous delivery robots on sidewalks by leveraging our experience on active sensing, multi-agent planning, and safe reinforcement learning. Dynamic and collaborative sensing mechanism for a multi-agent in the deep-learning era is challenging and has not been sufficiently studied. Particularly, the team aims to maximize the information gain given limited sensing, computation, and communication budget for each robot by collaboratively integrating the sensing systems between robots adaptive to different levels of telemetry communication bandwidths. By collaborative sensing with surrounding robots from the same company, the PDD fleet can achieve robust perception and decision making against environmental disturbances or non-stationarity.
The team plans to model the whole robust sensing, decision making, and communication procedure under the Constrained Markov Decision Processes framework and solve it using multi-agent reinforcement learning algorithms. The project will provide a novel solution to enhance the safety of autonomous delivery while reducing the cost, thus boosting the technologies and economy for the commonwealth. The research team will collaborate with Bosch Research and Technology Center in Pittsburgh and the city of Pittsburgh in technology development and testing.