On Valentine's Day in 2007, a human-sized robot called Dexter walked across a room in Mountain View. The journey was a little over six feet. It had taken nearly six years. Trevor Blackwell and his tiny Anybots team had already replaced the robot's heavy Doc Martens with Vans, built a more accurate gyroscope because commercial ones were not good enough, and watched innumerable attempts end in the embrace of a rolling safety cage. Then Dexter lifted one foot, placed it, caught its falling body and did it again. The room, for once, moved beneath the machine in the intended direction.
Blackwell's career makes better sense if you begin with that controlled fall. He is usually introduced with an enviable pile of nouns: programmer, engineer, roboticist, entrepreneur, investor. He co-founded Viaweb, the web-store company acquired by Yahoo in 1998. He co-founded Y Combinator in 2005. He founded the research company that developed Anybots. Yet his work is less a parade of institutions than a long fascination with feedback. Measure what is happening, correct before the error compounds, and accept that balance is a verb.
A shop window with a feedback loop
Blackwell grew up in Saskatoon, studied engineering at Carleton University and arrived at Harvard for doctoral work in computer science. His research concerned randomized network protocols, compiler optimizations and ways to measure the performance of complicated systems. The title of his dissertation, Applications of Randomness in System Performance Measurement, sounds far removed from a robot in skate shoes. It is actually a useful overture. Both problems ask what signal can reveal the truth about a system that refuses to sit still.
While at Harvard, he joined Paul Graham and Robert Morris in Viaweb. In the mid-1990s, letting an ordinary merchant create an online store through a browser was a considerable trick. Blackwell built merchant order processing and statistics systems, the parts that tell a shopkeeper what happened after someone clicked. Yahoo bought the company, and its software became Yahoo Store. Blackwell stayed through 2001.
He later joked that, back then, he thought the internet was about 80 percent finished. The estimate was premature, which is the sort of error that becomes charming only when followed by a good second act. He left Yahoo on a Friday and began Anybots the following Monday. Robotics offered an ocean of unfinished work and, inconveniently, several disciplines he did not yet know. He taught himself machining, mechanical design and control theory. A successful web founder had arranged to become a beginner again.
“Back in 2001, I thought the Internet was about 80 percent finished. That turned out to be a little premature.”Trevor Blackwell, recalling the start of Anybots
The six-year footstep
Dexter was the unreasonable project. Conventional demonstration robots could cross a prepared floor by replaying a carefully planned sequence. Dexter would balance dynamically, more like a person. Walking is continuous rescue. The body begins to fall, a foot arrives where it must, and the rescue starts all over again. On a human-sized machine, small errors grow with rude speed. By Blackwell's account, none of the available gyroscopes located the center of gravity precisely enough, so he made one.
The Blackwell loop
SENSE → ESTIMATE → ACT → REPEAT. Dexter's balance depended on seeing error early enough to move a foot before the error became a collapse.
Anybots then consisted of three people and a machine shop. Next door, or near enough to feel like the same experiment, the early Y Combinator ran its West Coast program. On Tuesday evenings founders could see the week's progress on Dexter. The arrangement had comic neatness: one room contained young companies trying not to fall over; the other contained a robot doing so literally.
When Dexter finally managed a few steps on February 13, the team removed the front of its cage to provide more room. The next day it walked far enough for the dimensions of the lab to become the limiting factor. One design detail survived because it is too good to discard: the Doc Martens had made the feet too heavy. Vans helped. Engineering history occasionally has excellent wardrobe notes.
The robot sent to the meeting
The research did not travel in a straight line from biped to product. Monty, another Anybots machine, had wheels, arms and the air of a useful laboratory creature. Its descendants became less humanoid and more practical. The turning point arrived with an expired passport. Blackwell was stuck at an airport in Canada while an important client reached his office. A speakerphone felt too meager. The team rigged its 200-pound robot, complete with arms and glowing eyes, so he could move around the meeting from afar.
What emerged was QB, a five-and-a-half-foot telepresence robot that weighed 35 pounds, balanced on two wheels and could be driven from a web browser. A camera, microphone, speakers and small screen gave the remote pilot a face, ears and a modest ability to wander. Anybots wrote its ambition on a whiteboard: “Make business travel obsolete.” At $15,000, QB was not a casual office accessory, but its argument was easy to grasp. A video call puts you on a wall. A mobile proxy lets you decide which wall to visit.
The engineering moved from laboratory spectacle to customer irritation. The machine needed to survive a car trunk. Its audio had to remain intelligible. Obstacle avoidance had to help without demanding the operator's attention. People wanted to see the pilot, so two-way video arrived. After many iterations, Blackwell described spending less time prescribing features and more time listening to users. His customers knew their workplaces; he knew how to build robots. The next correction came from comparing the two.
“The big manifesto we have written on our whiteboard is, ‘Make business travel obsolete.’”Trevor Blackwell on Anybots' aim
QB even acquired a small public life. It appeared at TED. Evernote chief executive Phil Libin used robots to move among offices in Mountain View, Tokyo and Moscow. A bystander filmed one rolling into a coffee shop. Blackwell punctured the implied marketing strategy with a simpler explanation: the team wanted a scone.
The accelerator beside the machine shop
Y Combinator began in 2005 with Blackwell, Graham, Jessica Livingston and Morris as its founders. Its batch model made small bets on young companies, put them in proximity and compressed months of learning into weekly conversations and a final presentation. Blackwell's name has never occupied as much public space as the accelerator he helped create. His own website resembles a workshop inventory rather than a monument: research papers, code, a balancing scooter, a startup growth calculator, a plagiarism detector and an Apple II screen mode for XScreenSaver.
That Apple II project is revealing. Blackwell did not merely reproduce green characters. He emulated the color and distortion of a late-1970s television set, then summed up the achievement with fond amusement for 40-by-24 pages of uppercase text. Accuracy can be practical or playful. In either case, it comes from noticing the behavior other people dismiss as noise.
Viaweb turns the browser into a shop-building tool.
Blackwell leaves Yahoo and starts his robotics work.
Y Combinator begins; Anybots shares the orbit.
Dexter's controlled fall becomes a walk.
Visibot asks how robot programming can become visual and legible.
Making the machinery visible
Blackwell eventually left the day-to-day roles that had made him familiar in Silicon Valley and is now based in Chipping Campden, England. His current project is Visibot, which he describes as a search for better ways to program robots. The available public software calls it a visual robot language. It is a fitting destination. His earliest company allowed merchants to construct something complicated through a browser. His telepresence robot hid motion-control machinery behind arrow keys. Visibot returns to the problem from the builder's side: how might a difficult machine reveal enough of itself to be programmed clearly?
The recurring idea is not that everything should be simple. Blackwell has repeatedly chosen work that is stubbornly complicated. The idea is that complexity needs a useful interface and an honest measurement. Viaweb showed a merchant the store. Dexter needed to know its center of gravity. QB needed to tell a distant operator where the room ended. Startups needed short loops between making and learning.
There is an appealing lack of ceremony to the whole record. Leave one field on Friday, begin the next on Monday. Replace the shoes. Build the sensor. Drive the robot to coffee. Publish the odd side project because someone else may enjoy it. The grand language can wait until the mechanism works.
A machine standing perfectly still is easy to admire and hard to trust. Blackwell's machines wobble, report the wobble and attempt a correction. So did the companies. So, one suspects, does the career. Its unity lies not in a tidy plan but in a practiced response to imbalance: look closely, move quickly and keep enough humor for the scone.