AI MATTERS, VOLUME 4, ISSUE 1 4(1) 2018 Artificial Intelligence in 2027 Maria Gini (University of Minnesota; [email protected]) Noa Agmon (Bar-Ilan University; [email protected]) Fausto Giunchiglia (University of Trento; [email protected]) Sven Koenig (University of Southern California; [email protected]) Kevin Leyton-Brown (University of British Columbia; [email protected]) DOI: 10.1145/3203247.3203251 Introduction Noa Agmon, Bar-Ilan University1 Every day we read in the scientific and popu- The discussion about the fourth industrial rev- lar press about advances in AI and how AI is olution, and the part of AI and robotics within changing our lives. Things are moving at a fast it, is wide. In the context of this revolution, au- pace, with no obvious end in sight. tonomous cars and other types of robots are expected to gain popularity and, among other What will AI be ten years from now? A tech- things, to take over human labor. While surveys nology so pervasive in our daily lives that we like “When will AI exceed human performance?” will no longer think about it? A dream that has (GSD+17) report that some researchers ex- failed to materialize? A mix of successes and pect robots to be capable of performing human failures still far from achieving its promises? tasks, such as running five kilometers, within At the 2017 International Joint Conference on ten years, this will probably take much longer Artificial Intelligence (IJCAI), Maria Gini chaired given the current state of robotic development. a panel to discuss “AI in 2027.” There were Today, the use of robots is generally limited to four panelists: Noa Agmon (Bar-Ilan Univer- three categories: non-critical tasks; settings sity, Israel), Fausto Giunchiglia (University of in which robots are semi-autonomous, tele- Trento, Italy), Sven Koenig (University of South- operated, or remote-controlled (namely, not ern California, US), and Kevin Leyton-Brown fully autonomous); and highly structured set- (University of British Columbia, Canada). Each tings in which uncertainties are minimal. Ex- of the panelists specializes in a different part amples of such settings include the Amazon of AI, so their visions span the field, providing Robotics warehouse robots, which work au- an exploration of possible futures. tonomously in a structured environment (the The panelists were asked to present their views warehouse), semi-autonomous drones oper- on possible futures, specifically addressing ated in military settings (usually follow a speci- what AI technologies they expected would be fied route autonomously, though operative de- in widespread use in 2027, what they thought cisions are made by human operators), robots would still show potential but not have become that perform cleaning tasks, which are consid- widely accepted, and what they expected the ered non-critical, Mars rovers, which operate AI research landscape to look like ten years semi-autonomously in unstructured environ- from now. ments, and more. When robots are required to operate fully autonomously in unstructured set- This article summarizes the main points that tings requiring them to handle unbounded un- each panelist made and their reflections on the certainties or completely unpredictable events, topics. The focus in each contribution is not they tend to fail. One of many examples is much on predicting the future but on bringing the Knightscope robot, which drove into a foun- up specific open problems in each subarea and tain on its first day of deployment as a security discuss how the current AI technologies could guard in Washington, D.C. be steered to address them. Rather than arguing about the ability of robots 1Acknowledgments: I would like to thank Gal Kaminka and David Sarne from Bar-Ilan University Copyright c 2018 by the author(s). for their helpful comments. 10 AI MATTERS, VOLUME 4, ISSUE 1 4(1) 2018 to outperform humans, and when this might back to the early 1980’s, when robots were happen, the following discussion examines the scarce and not autonomous. Research has challenges and opportunities that will influence progressed far beyond the deployment of such the development of intelligent robotics in the systems outside of labs. Improvements in the next ten years. reliability of robots will make it easier to deploy MRS in various applications, continuing and Dependence on hardware. As opposed to accelerating current successful trends (e.g., in the progress of AI, which relies mainly on al- warehouses and hospitals). This, in turn, will gorithmic development and benefits from pro- accelerate research on fully distributed, fully cessing improvements, progress in robotics is autonomous systems, which are beyond cur- also intimately tied to the capabilities of electro- rent capabilities. It is obvious that human-robot mechanics, physical sensors, and energy stor- interactions will be a major focus of research age and management. Whatever apocalyptic in the next ten years, as robots enter a greater or euphoric visions we have for working with number of unstructured environments in which robots, their realization is much more depen- humans operate. However, given the fore- dent on physical components than we, AI re- seen growth in the role of multi-robot systems, searchers and practitioners, tend to consider. human-MRS and multiple operator-single robot For example, most quad-copters, which are collaborations will likely see increased efforts. considered to be a basis for breakthrough ap- plications (such as home deliveries and emer- Increasing ties with other disciplines. A gency services), can only fly for 30 minutes or good example of large-scale fully distributed, so. Likewise, vacuum cleaners are limited in fully autonomous systems also raises an addi- the total area that they can cover before they tional trendthat of increasing ties with other have to be recharged. These energy concerns disciplines. Swarms of molecular robots radically impact the usefulness of robots in (nanobots), the size of which is measured in applications which are otherwise within reach nanometers, are becoming a reality in medical from a pure software perspective. applications (for example, targeted drug deliv- The good news is that the intimate connection ery). Trillions of such robots will be let loose between software and hardware works both in a patient’s body - the largest-scale MRS in ways. Just as modern SLAM algorithms (e.g., robotics history. The computation of interac- (DNC+01)) were able to overcome intrinsic sen- tions between different types of nanobots has sor limitations to create reliable and accurate an immense impact on the ease and duration of development of new treatments (WKKH+16; maps for navigation, advances in software can + overcome some of the limitations posed by KSSA 17). The capability to plan and rea- hardware. son about the interactions of these robots with each other and with the body requires deep AI influences on robotics. AI algorithms influ- collaboration between AI experts, biologists, ence robotics not only in compensating for and and chemists. improving the utilization of existing hardware capabilities, but also in enabling new tasks. Another example is reconfigurable robots, Progress in natural language processing (NLP) which can transform themselves into different and machine learning (used for chatbots, per- shapes, depending on the environment and sonal digital assistants, and surveillance, for in- task. Research on such systems will benefit stance) enables more natural forms of human- from close coordination with chemistry, physics, robot interaction with physical robots, and au- and biology, to take new findings into account. tonomous cars. However, such positive influ- ences are somewhat asymmetric: AI will influ- Increasing accessibility, lower entry bar- ence robotics more than robotics will influence rier, greater impact potential. A positive AI. A personal robot benefits from NLP more trend which I believe will continue to grow is the than NLP can benefit from the consideration lowering of the entry barrier into robotics prac- of multi-modal interactions (as in “talking with tice and research, at multiple levels. Research- your hands.”) grade robots for labs have seen dramatic de- creases in cost, and the common availability Growing role for multi-robot systems of 3D printing, cheap embedded computers (MRS). The academic research on MRS dates (Arduino, Raspberry Pi), as well as continued 11 AI MATTERS, VOLUME 4, ISSUE 1 4(1) 2018 push on STEM education will make develop- could not produce. ment of robots cheaper and easier than ever. The availability of common robot software mid- A lot of successful work in this area has been dleware, such as ROS, make it easier for re- done, see for instance (GT07; RKN16). How- searchers to focus their attention on bringing ever, the extent to which the KRR research their expertise to bear on specific components. could be improved by exploiting the research developed in ML, or dually, the extent to which Bottom line: More of the same (which is ML would really need to exploit any of the re- good!) Within the next ten years and be- sults developed in KRR is still unclear, at least yond, we will not see general-purpose robots. for two reasons. The first is that this integration That is, robots will still be dedicated to one is far from being trivial. It is a fact that these two task, for example delivery, cleaning, or surveil- approaches start from somewhat opposite as- lance. Progress in the development of intelli- sumptions, the first assuming that knowledge gent robotic systems will continue to focus on consists of a set of facts which are either true or excellence in the performance of specific tasks, false, with nothing in between, the second hav- and on the introduction of new tasks to new ing to deal with the issue that any fact learned types of robots. To some extent, this will make via ML will hardly ever be guaranteed to be robot use more popular.