Thoughts on Robot Autonomy
We are in the midst of tremendous advances in robotics. One of the most interesting aspects of recent progress in the field is the emphasis on increasing the level of robot autonomy. It is a longstanding quest of roboticists to build systems capable of ever-increasing degrees of self-sufficiency - whether these be robots in household settings, on the road, in the air, or in warehouses. For self-driving cars, there is even a systematic characterization of levels of autonomy with the highest (so called ‘full autonomy’) not requiring human attention. For other systems, autonomy may well mean careful and purposeful human engagement.
There has been explosive growth in machine learning technologies powered by advances in large neural networks. These have resulted in significant breakthroughs in computer vision and in natural language processing. They have not, however, resulted in commensurate breakthroughs in robotics. Why might this be? Hardly anyone questions that increased robot autonomy will depend crucially on robots being able to learn from experience. However, there is emerging evidence that progress in robotics will take more than simply applying techniques from machine learning to robots. In other words, robot learning is not the same as machine learning, or even machine learning applied to robotics - it is something else. With explosive growth and excitement has come a tendency to apply machine learning naively to robotics problems instead of a first principles approach to robot learning. I suspect the field will find a balance over time on this front, and will better understand the tradeoffs between learned and non-learned solutions to problems.
Autonomy in some settings may, somewhat paradoxically, depend on intelligent interaction. For some household robots to learn your habits efficiently it may be worthwhile for them to ask questions when they are first deployed, gradually tapering off such interaction over time. On the other hand, there are robots being built precisely for the purpose of interaction with humans, often to provide assistance with aspects of everyday life and well-being - both psychological and social. Quite opposite to tapering off their interaction, these robots will need to interact with people in meaningful ways over extended timescales, and the quality and richness of that interaction will make all the difference to their utility and adoption. This certainly doesn’t mean that such robots won’t be autonomous - quite the opposite! Humans interact substantially with each other via language and developments in large language models have spurred recent research in robot-human language-based interaction. Large language models have proved surprisingly versatile as a basis for tasks they were not trained on, thus leading to the exciting idea that they, or variants, may serve as so-called ‘foundation models’ for ‘downstream tasks.’ Is the quest for autonomous robots simply a ‘downstream task?’ Time will tell.
Autonomous Robots publishes original papers in all aspects of robot autonomy with a particular emphasis on experimental work with physical robots. The journal advocates the idea that embodiment is fundamental to making progress in robotics, particularly when it comes to robots that learn from experience and interact with humans - both key drivers in the quest for autonomy. With so much going on in the field, this is an exciting time to be a roboticist!
Gaurav S. Sukhatme is Professor of Computer Science and Electrical and Computer Engineering at the University of Southern California (USC). He holds the Fletcher Jones endowed chair in Computer Science. He is currently an Amazon Scholar. Sukhatme has served as the Executive Vice Dean at the USC Viterbi School of Engineering and the chair of the Computer Science department. He directs the USC Robotic Embedded Systems Laboratory, which he founded in 2000. His primary research interests are in multi-robot systems and robot networks. He is a Fellow of the AAAI and IEEE. He is a founder of the Robotics Science and Systems conference and was program chair of the 2008 ICRA and 2011 IROS. He is the Editor-in-Chief of Autonomous Robots and has served as Associate Editor of the IEEE TRO, the IEEE TMC, and on the editorial board of IEEE Pervasive Computing.