Autonomy in Transportation: Emerging Challenges in multi-agent planning and control

Overview

With the advent of self-driving cars and delivery robots, automation promises to revolutionize the transportation sector in our modern society. Unfortunately, autonomous vehicles and robots in real-world traffic often cause collisions and experience gridlock, raising doubts about their ability to safely and efficiently interact with surrounding vehicles and pedestrians. To address these issues, researchers across control, robotics, and machine learning have developed a wide range of methods to improve the reliability of autonomous vehicles and robots during their deployment. Such techniques include implicit scene representations for high-accuracy autonomous perception, to dynamic games and multi-agent reinforcement learning for interactive prediction and motion planning, to distributionally robust control or barrier function or model predictive control-based methods for designing safe control strategies. However, it is often unclear whether the inefficient or unsafe behaviors exhibited by autonomous agents should be attributed to failures of the perception, estimation, or multi-agent prediction and planning stack, and could thus be mitigated by designing more reliable methods for the corresponding portion of the autonomy pipeline. Moreover, since the design and deployment of these methods in real-world interactive robotics applications are nascent, the relative merits and limitations of each method remain unclear. Thus, to develop holistic perspectives of state-of-the-art approaches for assured multi-agent autonomy, this workshop will address the question: How can we design prediction and planning paradigms to guarantee safe and performant multi-agent interactions in emerging autonomous transportation platforms?

Our workshop aims to tackle the above question by examining various perspectives from the fields of controls and robotics on emerging and urgent challenges in autonomous transportation. Concretely, we will bring together speakers representing expertise in control theory, dynamic games, and machine learning. We will hear from Dr. Sarah Li (GT Aero), who will present her work on stochastic reach-avoid problems within the context of Markov Potential Games; Dr. David Fridovich-Keil and Dr. Jingqi Li (UT Austin Aero), who will present their work on the design of efficient algorithms to solve multi-agent, non-cooperative dynamic games; Dr. Samuel Coogan (GT ECE), who will present his work on control barrier function-based methods for human-autonomy teaming; and Dr. Chih-Yuan Chiu (GT ECE), who will present an organizer talk on robust game-theoretic motion planning methods for safe multi-agent navigation under intent and dynamics uncertainty. We currently also aim to invite one or two additional researchers in multi-agent autonomy to speak at our workshop, pending their confirmation. Moreover, we will solicit a group of students and postdocs to present posters on a broad range of topics, aiming to seed valuable, cross-disciplinary discussions. Alongside these poster presentations, the broad expertise of our invited speakers will highlight the main challenges in deploying autonomous vehicles in societal-scale transportation systems and the proposed solutions to address these challenges.

More information regarding our invited talks can be found at the Invited Presentations tab.

We have divided our workshop into three session types, each of which allows for a different mixture of voices to be heard:

• Invited Talks - These sessions will allow attendees to learn about recent results and current trends from experts in the research areas of multi-agent motion planning in the context of transportation and autonomous vehicles.

• Interactive Poster Session - This session will showcase the contributed work of young researchers and will provide opportunities for brief talks and informal discussions between participants and speakers. Contributed poster papers will undergo a limited review process, and the presenters will be showcased on the conference website. Presented works may appear elsewhere in the conference program, but need not.

• Panel Discussion - This moderated discussion will include our invited speakers. The moderator will prepare several questions for each panelist. Attendees may also ask questions and participate in the discussion, which the moderator will select for the panelists. Tentative discussion topics include:

• Q1: Which parts of the autonomy stack can be performed by AI and which should not? How can we ensure that AI-empowered multi-agent motion planning algorithms prescribe reliably safe behaviors despite unforeseen AI failures?

• Q2: In the field of multi-vehicle autonomy, what are some overutilized or underutilized methods for model, algorithm, or experiment design?

• Q3: What are the biggest safety gaps in currently deployed and emerging intelligent transportation systems? From an engineering perspective what are the most promising approaches to address them? From a consumer perspective, what milestones need to be reached?

• Q4: Multiple approaches to autonomous vehicle fleet operation have now been deployed in real-world traffic, including Waymo fleet of individual rides, Tesla Robo-taxi consumer owned autonomous vehicles, Zoox autonomous pods or pools. How does each approach align with your opinion/direction for autonomous vehicle research and development?

• Q5: What forms of vehicle-to-vehicle communication and vehicle-to-infrastructure communication are reasonable and valuable?

• Q6: There have been several high-profile incidents, such as all Waymo cars stopping at the same time in busy traffic, or overcrowding a single charging station. How can we leverage inter-vehicle communication or other methods to avoid similar interaction-related incidents in the future?

Areas of interest: 

Autonomous Vehicles, Safety, Multi-Agent Reinforcement Learning, Route Planning and Prediction, Human-Autonomy Teaming

Techniques include: 

Hybrid Systems and Control, Game Theory, Barrier Functions, Markov Games, Dynamic Games, Optimization, Reinforcement Learning