In March of 2019, Boston Dynamics released yet another amazing video showcasing the box stacking capabilities of their whegged (wheel-legged) robot. The robot zooms around a warehouse, picking up and placing objects on its two wheels.
As Wired put it, this represents Boston Dynamics being “back at it”, with ‘it’ referring to their periodic release of videos showcasing their robots’ performing amazing feats. The previous such video came out in October 2018 and showcased the parkour skills of their humanoid robot Atlas. The robot can be seen soaring over logs and leaping onto platforms – actions which require extreme energy and strength for a hefty robot to perform (Atlas weighs up to 150kg). Nevertheless, the video shows Atlas navigating these terrains with remarkable ease:
Marc Raibert, former Carnegie Mellon and MIT professor and director of the Leg Lab, developed the first self-balancing hopping robots. In 1992, he went on to found Boston Dynamics. Since then, you may have heard of them through many of their other projects. Some favorites include BigDog, capable of traversing rough terrains; SandFlea, which stores enough energy to jump onto buildings; SpotMini, which brings smooth moves to the dance floor; and Atlas, the humanoid robot that can drive its own vehicle.
In particular, Atlas has demonstrated many impressive actions throughout the years. Its capabilities include backflips, opening and holding doors, washing dishes, trail running, and lifting boxes, and more:
Atlas initially triggered a dystopian response from viewers and media alike. Many of these headlines appear to be in jest. However, there also seems to be serious concern from some viewers about the true implications of Atlas’ current capabilities, and some of the following don’t help:
Fast Company greeted the new warehouse robot with the headline “Those frightening Boston Dynamics robots will also take your warehouse job”.
The Telegraph published a summary of Atlas’ backflip video calling the robot a “terrifying humanoid.”
At the time of the reveal of the original Handle robot, The Inquierer called it a new robot nightmare.
Many of the “humorous” headlines referenced the pop-culture concept of a “robot uprising.”
Gizmodo warned of a “shockingly nimble” robot uprising.
MarketWatch advised readers to “not bother running away.”
Mashable has a top 5 of “times Boston Dynamics robots scared the hell out of us”.
This, in turn, has led to some concerns about misunderstandings that may arise from Boston Dynamics’ strategy of releasing these videos without much further explanation:
- Researchers and other experts are more worried about people’s raised expectations about what robots really can do than by the robots themselves. NVIDIA’s director of research describes her experiences with the general public, stating that many believe the Boston Dynamics dogs are more intelligent than real dogs.
- Filip Piekniewski, an AI researcher, likewise expressed that the videos do not mean that Boston Dynamics’ robots are anywhere near as intelligent as dogs:
- Jack clark, a journalist, stated that Boston Dynamics’ press strategy was simply to release videos and avoid any questions from the press:
It’s hard to get a sense for the general public’s perception of the videos, but just sampling several tweets shows some are genuinly concerned about the ‘robot apocalypse’ scenario:
Of course, these represent outliers in the public opinion. Still, reactions along the lines of ‘impressive, but also vaguelly offputting’ are common whenever new Boston Dynamics videos come out:
Believing that these videos represent robotics with advanced autonomous capability is problematic, since that is far from the truth. There are substantial differences between a robot that autonomously performs these actions versus one that is preprogrammed to do so with the help of external sensors and computation, and the lack of transparency makes it unclear what Atlas and friends are capable of. Though the details are not clear, it appears to be the case that Boston Dynamics’ technology does not involve learning or advanced cognition and is based on human-defined methods for robotic control:
Fortunately, multiple articles have tackled the subject of Boston Dynamics videos. A Wired article does a good job explaining how to watch and interpret Boston Dynamic’s: focus on the large margins of error (there’s always a chance in the demos that things can go wrong, but what you’re seeing is the successful trial), partial autonomy (the systems don’t operate entirely on their own and require human guidance or pre-determined instructions), and our human instinct to anthropomorphize and connect to biological systems (humans tend to see the dog- and human-looking robots and assume many advanced capabilities that the biological versions already have, such as perception and emotion). Additional articles tackling the subject include Boston Dynamics’ scary robot videos: Are they for real?, Why Boston Dynamics’ backflipping borg shouldn’t scare you, You’re Expecting Too Much Out of Boston Dynamics’ Robots, and The Fantasy Robots Of Boston Dynamics. As the last piece states:
“The achievements of Boston Dynamics in robot movement on two and four legs are undeniable and sometimes breathtaking. That’s where its real contribution lies. But its videos play on widespread, potentially dangerous fantasies of autonomous robotic armies without human casualties (on our side, that is), old people left in the care of machines, and a near future when robots do all the work. It’s high time we freed ourselves from the techno-marketing hype and began to see AI and robotics for the fascinating, although limited, developments they really are.”
Alongside their video, Boston Dynamics released a caption with their video, stating:
Handle is a mobile manipulation robot designed for logistics. Handle autonomously performs mixed SKU pallet building and depalletizing after initialization and localizing against the pallets. The on-board vision system on Handle tracks the marked pallets for navigation and finds individual boxes for grasping and placing.
When Handle places a boxes onto a pallet, it uses force control to nestle each box up against its neighbors. The boxes used in the video weigh about 5 Kg (11 lbs), but the robot is designed to handle boxes up to (15 Kg) (33 lb). This version of Handle works with pallets that are 1.2 m deep and 1.7 m tall (48 inches deep and 68 inches tall).
While it’s nice to see this description being released alongside their video, it would be helpful if Boston Dynamics would reveal more about their methods. The about page of their website does not contain more details, and their robot-specific pages explain slightly more but do little to dispell misunderstanding. The lack of concrete details is in part responsible for the rampant speculation online, and in fact the company seems to provoke and enjoy the speculation by posting viral, vague videos. This helps to create interest around Boston Dynamics’ projects, and their incentives to be as opaque as possible are almost understandable: secrecy insures that competitors cannot copy their achievements, strikes the public’s imagination, and leaves everyone in the dark about the weaknesses of their technology. Yet the dearth of information is not all benefits: it creates unrealistic expectations for real world performance of the robots, and provokes the ire of less naive observers.
An article on the potential dark future of AI spells out the parallels of SpotMini from Boston Dynamics to a Black Mirror episode containing weaponized guard dogs that look a little too similar for comfort and more importantly goes into detail about what exactly people fear. The list includes aspects such as the incomprehensible, black-box decision-making process of robots, particularly when it comes to significant decisions that may infringe upon international human rights law and international policing standards. While Boston Dynamics does not explicitly make any egregious claims on how advanced their technology is (unlike Sophia), their non-transparency does not help assuage these fears.
These threads and discussions prompt some of the same questions that we came up with while watching the videos:
Are the actions and motions pre-programmed? Does the sequence of the obstacles impact performance?
Are Atlas’s decision-making process and motion planning occurring in real-time?
Is the computation being done on-board? How much information is embedded within the visual markers?
How much information does Atlas have about its environment? What knowledge is transferable to other environments?
Many of these questions represent open challenges in robotics and autonomy; some transparency would be helpful to better gauge which problems Boston Dynamics has solved. In addition, there is a gap between systems that can perform a few of these individual tasks in structured environments and those that can autonomously navigate in unstructured terrain. To be fair, what Boston Dynamics has accomplished is still very impressive. Their robots are capable of many extremely dynamic tasks and appear to be incredibly robust to disturbances in their environment. However, transparency would help both researchers and general public alike to better understand which features are realistic and which are fantasy.
In addition, there is genuine concern from the general public and researchers alike about the use of robots as autonomous war weapons. Boston Dynamics makes no secret of their interest for military applications, and the misguided headlines about a “robot uprising” against humans obfuscate the more mundane prospect of humans killing other humans with more ease than ever, using robots.
Atlas is a fantastic case study of a robot that demonstrates the significance of distinguishing between one-off demonstrations and generalizable techniques that are usable in various contexts. The physics and mechanics of each capability are incredibly impressive, but we are very far from the cries of a robot uprising. For comparison, trickshot specialists produce jaw-dropping highlight reels, but their skills don’t always translate directly to the level of professional athletes who must seamlessly weave multiple skills together simultaneously. Real-world environments are very complex, and systems need to be robust to uncertainty in their environments.