Not long ago, using the term artificial intelligence to describe actionable decision-making by machines could be controversial depending on your background, especially in the world of academia where many saw it more as a marketing buzzword than an actual technology. In a practical sense they weren’t wrong, and even today, you should be skeptical when you see a company use it in the context of one of their products. However, as more products demonstrate behavior that resembles actual intelligence, and the pace of progress in machine learning accelerates, it becomes hard not to reconsider our definitions of what the limitations of machine intelligence are, and how it will impact our lives.
With the growing global demand for machines that can perform labor, and an understanding of how to build these systems, we can start to meet that demand. Robots are now learning to take on increasingly open-ended tasks without the requirement of reprogramming for each new instruction. What used to be a long lead-time in deploying new systems is disappearing as robots can reuse not just hardware, but prior knowledge when taking on new tasks. This means we’ll see more of them moving into factories to boost efficiency, filling in where labor shortages exist, and in places where human workers can’t go because of hazardous conditions.
Technologically, the last 30 years or so have been shaped by advancements in computation, disseminating the effects of Moore’s Law to fuel the supercharged growth of the internet and support large-scale systems for the digital industries. The ability to build machines that can make decisions independent of human operators is, of course, a direct result of this progress, but the shifting momentum toward intelligent automation on specialized systems will bring about prerequisite changes needed to deliver at scale.
How Machine Learning Has Been Shaped By the Web
While humanoid robots often unfairly get the majority of attention as the embodiment of all this progress — mostly because they make for eye-catching banner graphics — the progress in A.I. up to this point has disproportionately served an information economy based on the web. Where there is an objective to optimize, ecosystems of hardware and software have sprung up in support, conceding to Darwinian selection as a natural consequence. Advancements over the beginning of the 21st century have followed profit in e-commerce, social media and search engines, and have resulted in impressive systems that can optimize ad placement to influence people’s behavior at scale.
Over this same period, robotic systems with embedded intelligence have made progress, but on a smaller scale. It’s worth pointing out that, in many respects, these systems are inherently more difficult to build with reliability because they operate in the real world — a world with less regularity and more unpredictable consequences than the carefully-designed frameworks of the digital world. Software and hardware become equally important in this setting and must be both robust and performant. Given the challenges, it’s not surprising to consider that factory automation is still largely driven by mechanical design, and supplemented by human workers which perform tasks that are often repetitive, but difficult to automate.
Automation may be challenging, but it’s not the case that research can’t solve a lot of these problems — just that they became a possibility only recently. There are certainly quite a few reasons for this. Some have to do with hardware and computational advancements that have sprung from technological prosperity, and others revolve around data strategies and deep learning. Viewing the systems landscape more holistically, another interesting theory has emerged around the over-optimization of the platforms that benefited the web industries. In “The Hardware Lottery”, Sara Hooker looked at machine learning advancements through the lens of hardware and software availability. She postulated that research ideas may succeed or fail as a result of domain specific hardware intended and optimized for orthogonal industries.
The limitations we see in some areas of A.I. research may be due to this over-committing (or over-fitting) to one problem, which subsequently limits progress in solving another, effectively making them more costly to work on. Robots, indeed, are heavily dependent on specialized hardware and software to operate, so it’s not unreasonable to think that this could be another historical reason why intelligent robots aren’t as ubiquitous as we might expect them to be. This remains a monumental opportunity for the industry to develop systems that inherently support machine learning research and aren’t just an afterthought. Viewing robots from this perspective is an argument for why both hardware and software must coevolve to make significant forward progress.
Working to Bring AI Advancements to Robotics and Automation
By now, it’s becoming apparent that we have reached an inflection point. Machine learning for robotics has progressed despite a dearth of proper resources. Many of the same methods and theoretical work that have led to success in strategy games and computational biology are being applied to applications that can also enable greater autonomy in machines. Self-driving cars, unmanned aerial vehicles and spacecraft are still relatively in their infancy, but are making remarkable progress and in many capacities are able to outperform humans in head-to-head competition. Robots are demonstrating increased mobility, dexterity and flexibility when encountering new challenges that used to be the purview of special effect artists.
The difference between now and a few years ago is that we are taking the lessons of the internet and are beginning to see the possibilities of applying data at scale to robotics. All the orthogonal work that went into building the hivemind known as the world-wide web is being harnessed to enable more capable machines. You might say that we’re seeing some combination of the law of unintended consequences colliding with human ingenuity. The same systems that connect ideas from around the world and deploy predictive models around movie preferences, fashion trends and social networks are making it possible to connect robots in diverse scenarios all around the world. The same hardware that runs online multiplayer games, creating open-world environments with immersive graphics, process sensory inputs for autonomous machines in the form of deep learning models. At this point, the real promise of applying machine learning to robotics isn’t teaching a single robot to learn for itself, but to aggregate experience from a vast network of robots so that they can improve at scale.
I work with a talented team of engineers at RIOS who come from diverse backgrounds to build robots with this idea in mind. Hardware, skills and behaviors should be transferable across platforms when possible, and each deployed system should be able to share what it has learned with other systems. From a hardware perspective, we build modular systems that are a large-scale equivalent of LEGO sets that can be reassembled on demand. From a software perspective, you can think of what we do as storing knowledge rather than just data to reduce the need for retraining. The result is a class of robots that can do a variety of tasks and address new challenges with less development time. By building distributed robots that continuously learn from both their environment and the collective experience of others, we can help push intelligent robotics forward at scale much in the same way that the information economy benefited from the web.
- Yevgen Chebotar, Ankur Handa, Viktor Makoviychuk, Miles Macklin, Jan Issac, Nathan Ratliff, Dieter Fox: “Closing the Sim-to-Real Loop: Adapting Simulation Randomization with Real World Experience”, 2018; arXiv:1810.05687.
- Sara Hooker: “The Hardware Lottery”, 2020; arXiv:2009.06489.
- Neha Soni, Enakshi Khular Sharma, Narotam Singh, Amita Kapoor: “Impact of Artificial Intelligence on Businesses: from Research, Innovation, Market Deployment to Future Shifts in Business Models”, 2019; arXiv:1905.02092.