Are Robots Smart?
I’d argue that they aren’t. But let me provide you some context first.
Robotic technology is advancing at an impressively rapid rate. Mechanical and electronic components are becoming more efficient and powerful every year, allowing for more versatile and elegant robotic designs and applications. And on the software end, the artificial intelligence revolution has paved the way for machine learning and computer vision technologies, unlocking new ways for robots to interact with and interpret the world (OpenAI — Solving a Rubik’s Cube with a Robot Hand, Google AI — Quadruped Robot Learns to Imitate a Dog).
Boston Dynamics has emerged at the forefront of this robotic revolution, combining cutting-edge software with various quadrupedal and bipedal robots. I’m sure you are familiar with their designs, since almost all of their recent videos have gone viral. But if you haven’t watched their recent robot parkour video, check it out!
Impressive, right?
The speed at which Boston Dynamics’ robots are improving is frankly mind-boggling. Going from having trouble balancing to effortlessly vaulting over obstacles in a span of five years is a testament to the talent at the company and the speed at which our technology is improving.
If you noticed the background of the parkour video, you may have noticed the large barriers surrounding the robots’ operation area. In many industrial settings, you will see similar features — robotic arms surrounded by fences covered with “caution” and “danger” symbols. These barriers exist to protect humans from robots, because unfortunately these bundles of sensors and mechanical actuators do not account for the fact that they are seriously fast and powerful. These robots are hazardous not because they are too artificially intelligent, but rather because they are oblivious and not smart enough for a world inhabited by fleshy and fragile humans.
To put it more generally, most industrial robots are contextually unaware. They are unmatched in their ability to perform repetitive and precise actions, but falter when faced with unexpected contexts and new environments. Naturally, this severely limits their applicabilities and use cases. Despite the extraordinary hardware powering these machines, humans and robots still seem to inhabit two different worlds.
So how can we help robots overcome these (figurative and literal) barriers?
Collaborative Robotics
Collaborative robots, or “Cobots” for short, seek to bridge the gap between the human and robot environment. These robots incorporate features that optimize safety and familiarity to humans. For example, Rethink Robotics’ “Sawyer”, a popular cobot, features a screen displaying a pair of expressive cartoon eyes. The Willow Garage “PR2” takes this one step further by having an approximately humanoid form complete with two arms and an actuating head full of sensors and cameras. Other, more subtle hardware design choices for cobots include the usage of lightweight or soft materials, organic-inspired geometries such as rounded edges and naturalistic curves, as well as lower operating velocities and torques.
Research has shown that humans are more effective at interacting with robots that mimic human features or behaviour. For example, an experiment conducted in 2014 with the aforementioned “PR2” tested a basic handover scenario where the robot would hand a water bottle to a human recipient. The researchers found that participants reached for the bottle faster and more confidently when the robot expressed human-like nonverbal gaze cues such as making eye contact with the participant [1].
In nearly all robotic tasks, the robot needs to be programmed and taught how to execute the desired action. Currently, the most widely accepted and commonly used standard to accomplish this is through the usage of teaching pendants, which are essentially large remote controls reminiscent of graphing calculators or GameBoys. Although these are effective tools for communicating basic commands and trajectories without having to write code, they are not very intuitive and require technical training in order to operate effectively [2]. As automation continues to gain prevalence in industrial and home settings, we will need an alternative teaching method that is friendlier to those with non-technical backgrounds. Enter Learning from Demonstration.
Learning from Demonstration
Made possible by the rise of cobots that are safer and more intuitive to interact with, researchers have been exploring methods for robot teaching that are more kinaesthetic and hands-on. Learning from Demonstration (LfD) is one of those methods, in which humans manually position and maneuver the robot to teach actions [3]. Using LfD, teachers no longer need to understand the intricacies of robotic programming.
So how does LfD work?
Imagine you want to teach your future robot assistant how to pour a glass of juice. Since cups come in a variety of sizes and shapes and everyone’s kitchen layout is different, there isn’t a simple “one size fits all” set of commands that will be applicable in every user’s home space. Traditional teaching methods would not work very well in this context.
Using LfD, the process is trivialized. You simply move the robot through the desired motion, have it pick up the pitcher and navigate to the cup, then subsequently pour the juice. Move the pitcher and cup to new positions on your counter, then guide the robot again. After a few of these demonstrations, the task will be fully learned! The robot will then be able to pour you a glass of fresh juice regardless of where on the counter the cup and pitcher are located.
A wide range of tasks can be taught using this method, since the performance of the robot is only really limited by its hardware and the user’s teaching ability. At the same time, the user is spared from needing to interact with complicated teaching pendants or, God forbid, write code. Additionally, since the LfD process mimics how humans learn, we have an innate intuition for the teaching process. A notable disadvantage of LfD is that it has low accuracy; since the teacher is manually positioning the robot, it is difficult to be precise when aiming for a specific point or coordinate. This is not a very significant concern in home environments, but is a lot more significant in industrial applications where precision and repeatability is paramount.
Behind the scenes, Learning from Demonstration is enabled by machine learning techniques such as Gaussian Mixture Models (GMM) [4]. This summer, I have been investigating these models to find ways to improve the LfD process, particularly for those that are more inexperienced or have non-technical backgrounds. These ways include implementing an uncertainty metric to inform the user where the robot is not yet confident in performing the task (and thus request a demonstration from that uncertain state), as well as creating real-time visualizations to communicate the learning process. Through this research experience, I have learned a lot about collaborative robotics, human-robot interaction and kinaesthetic teaching techniques like LfD, as well as a bit about human psychology and how we interface with technology. Armed with this knowledge and wanting to pursue this field further, I am excited to take on my next internship — building autonomous home robots at a bay-area startup called Matician.
So, are robots smart?
…Well let’s face it, it’s a bit of a silly question. “Smart” is a completely relative term. But based on my experiences this summer, I define robotic “smartness” as the ability for robots to coexist with humans in a human-centric society.
And going by that definition, I’d argue no. Because despite the huge advancements in hardware, AI, and collaborative robotics, we still have plenty of barriers to tear down.
I would like to extend a huge thank you to Dr. Carol Jaeger and Dean James Olson for supporting the Cansbridge fellowship at UBC. In addition, I would like to thank Mitacs for sponsoring this year’s cohort, this work was supported by Mitacs through the Mitacs Accelerate program.
References
[1] A. Moon et al., “Meet me where i’m gazing: How shared attention gaze affects human-robot handover timing,” in Proceedings of the 2014 ACM/IEEE international conference on Human-robot interaction, 2014.
[2] C. Heyer, “Human-robot interaction and future industrial robotics applications,” 2010 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2010, pp. 4749–4754, doi: 10.1109/IROS.2010.5651294.
[3] A. Billard, S. Calinon, R. Dillmann and S. Schaal, “Robot programming by demonstration”, Handbook of Robotics, Springer, 2008.
[4] Calinon, S. A tutorial on task-parameterized movement learning and retrieval. Intel Serv Robotics 9, pp. 1–29, 2016. doi: 10.1007/s11370–015–0187–9
