The seminar “Bringing AI Up To Speed” discusses the challenges of developing AI for real-world physical tasks, especially autonomous driving, highlighting the complexity and unpredictability that make it much harder than closed systems like chess. It covers research on measuring and improving AV safety, the use of autonomous racing as a testbed for physical intelligence, and emphasizes the need for robust real-world testing to achieve human-level performance in robotics and self-driving cars.
The seminar, “Bringing AI Up To Speed,” explores the challenges and progress in developing artificial intelligence for real-world physical tasks, with a particular focus on autonomous driving and robotics. The speaker begins by contrasting the apparent intelligence required for tasks like chess with the complexity of driving, noting that while AI has long surpassed humans in closed systems like chess, it still struggles with open, unpredictable environments such as roadways. This is due to the immense “coverage complexity” of real-world driving, where countless variables—objects, environments, and interactions—must be accounted for, making it a far more difficult problem for AI than structured games.
A significant portion of the talk addresses the current limitations of autonomous vehicles (AVs). Despite impressive advancements, AVs still require human oversight and intervention, often failing in unpredictable or nuanced scenarios that good human drivers handle with ease. The speaker emphasizes that while AI has made strides in open systems like language and vision, the leap to “physical intelligence”—where AI must understand and interact with the real world, experiencing the consequences of its actions—is a much greater challenge. This gap is evident in the persistent need for human “babysitting” of self-driving cars and the difficulty of simulating the full range of real-world driving situations.
To address these challenges, the speaker’s research group has focused on developing methods to empirically measure and improve AV safety. One approach involves comparing the safety of different AV systems by automatically mining and analyzing similar driving scenarios from large datasets, using scenario description languages, trajectory analysis, and even language models to assess similarity. Another line of research uses adversarial reinforcement learning in simulation to automatically generate scenarios where AVs are likely to fail, thereby identifying weaknesses and improving robustness through repeated testing and hardening.
The seminar then shifts to the use of autonomous racing as both a testbed and a catalyst for advancing physical AI. Drawing parallels to the early days of motorsport as a proving ground for automotive technology, the speaker describes the development of scaled-down autonomous race cars (F110th) and the organization of global competitions to foster research and innovation. This work eventually scaled up to full-size autonomous race cars, culminating in participation and success in the Indy Autonomous Challenge, where the team achieved record-breaking speeds and demonstrated advanced planning, control, and adaptability in high-stress, high-speed environments.
Throughout the talk, the speaker highlights the importance of real-world testing, robust engineering, and interdisciplinary teamwork in pushing the boundaries of AI for physical tasks. The ultimate goal is to develop AI drivers that can match or surpass the skills of expert human drivers, not just in controlled environments but in the unpredictable, dynamic real world. The seminar concludes with reflections on the ongoing journey toward human-level physical intelligence in AI, the lessons learned from both successes and failures, and the exciting future of autonomous racing and robotics research.