There is an odd continuity between a palm-sized robot making eye contact and an excavator cutting a grade within centimeters. Both are machines trying to behave convincingly in a world designed for people. Both must sense, decide, and move. The difference is what happens when they get it wrong. Cozmo, the expressive robot Boris Sofman helped create at Anki, might tumble off a table. A construction machine can swing tons of steel across a live job site. Sofman's career has widened that gap one machine at a time.
His route into artificial intelligence began with a world inside a computer. As a boy, Sofman played Civilization, the strategy game in which research, resources, armies, and cities all compete for attention. He was fascinated by the software behind the coordination. The AI was imperfect, but the scale of the problem held him. Decades later, the objects have changed from pixels to machines, while the question remains recognizable: how can an intelligent system make useful decisions amid constant change?
Sofman was born in Moscow in the Soviet Union. His parents had tried to leave for more than a decade before the family was permitted to emigrate in 1989, when he was six. They passed through Europe, lived for a time in Brooklyn, and eventually settled near Dallas, where his father worked in telecommunications and large-scale optimization. Mathematics, engineering, and systems thinking were part of the household atmosphere. So was the lived understanding that a new environment changes the choices available to you.
A robot learns the room
At Carnegie Mellon, Sofman studied electrical and computer engineering alongside computer science. He kept a 4.0 undergraduate record, captained the varsity tennis team, and stayed for a PhD at the Robotics Institute. His thesis, completed in 2010, examined online learning techniques that could improve robot navigation in unfamiliar domains. The title sounds academic. The underlying problem would follow him everywhere: a machine leaves the neat boundaries of the lab and meets a place it has never seen before.
His advisers, Drew Bagnell and Tony Stentz, supported an entrepreneurial experiment growing beside the research. With fellow graduate students Mark Palatucci and Hanns Tappeiner, Sofman founded Anki in 2010. The founders had worked around space, agricultural, and industrial robotics. They chose entertainment as the doorway into ordinary homes. A racing game could make serious technology legible. It was a pitch you could understand by watching two little cars fight for a corner.
2010-2019
2019-2024
2024-now
Anki Drive debuted during Apple's 2013 Worldwide Developers Conference keynote, an unusual stage for a robotics startup that had not yet put a product on sale. The phone was not simply a remote. It acted more like a conductor, coordinating physical cars that could perceive the track, compete, and make decisions. Then came Cozmo, a tiny tracked robot with animated blue eyes, a lift for a nose, and behavior shaped with help from people who understood screen character. Sofman argued that perfection was boring. A believable robot needed quirks, pauses, and the capacity to surprise.
“Construction machines don't just navigate the world - they sculpt it with centimeter-level precision.”Boris Sofman
Anki shipped more than 3.5 million devices and approached $100 million in annual revenue, but hardware success did not remove financing risk. The company closed in 2019 after failing to secure a critical round. There is no tidy moral in that ending. A product can delight customers, a team can solve difficult engineering problems, and the business can still run out of road. What survived was a group of people who knew how to make embodied intelligence behave.
Thirteen engineers change lanes
A few months after Anki shut down, Waymo hired Sofman and 12 former colleagues. He took charge of engineering for commercial trucking and ultimately served as senior director of engineering and head of trucking. The jump from toy cars to Class 8 trucks became an irresistible joke among friends, but it was also a serious transfer of craft. Perception, planning, machine learning, simulation, and safety all grew more consequential at highway speed.
Waymo offered a view of autonomy at a scale few teams had seen. Sofman watched machine-learning-driven systems overtake hand-built collections of rules. Public roads produce too much variation for engineers to anticipate every scene. A system had to learn patterns from enormous data sets and generalize across vehicles and cities. By the time he left in March 2024, autonomous driving had become less of a laboratory promise and more of a daily service.
The next problem arrived through scale in another sense. America wanted more housing, factories, energy projects, and data centers. Construction companies had the work but not always the people to do it. Sofman and fellow Waymo alumni Kevin Peterson, Tom Eliaz, and Ajay Gummalla formed Bedrock Robotics around a practical idea: give contractors more capacity by making their existing heavy equipment autonomous.
The aftermarket future
Bedrock does not begin by asking a contractor to replace an excavator that may have cost $500,000, and sometimes much more. Its Operator system is a reversible retrofit: cameras, lidar, positioning, computing, controls, and software are installed on machines already in the fleet. Manual operation remains available. This product choice is also a lesson in adoption. The fastest route to a new industrial system may be through equipment whose economics are already understood.
Excavation turns autonomy into a moving physics problem. A road vehicle tries to understand a world that mostly stays in place. An excavator changes the topology with every scoop. Soil differs. Weather changes traction. Loads shift. The machine has to follow a site plan, avoid people and equipment, control a hydraulic body, and leave behind the intended terrain. The data is not merely about where to go. It captures how an expert operator gets work done.
Fit the future onto the installed base. A reversible retrofit lowers the capital and trust costs of trying something new.
Measure work in the customer's unit. Cubic yards moved says more to a contractor than an abstract autonomy score.
Let experts shape the data. Operator technique contains nuance that hard-coded rules cannot capture cleanly.
Preserve the manual path during transition. New capability feels safer when the old control remains within reach.
Bedrock emerged from stealth in July 2025 with $80 million in seed and Series A funding and machines working with construction partners across four states. In November, its supervised autonomous excavators worked alongside Sundt Construction on a 130-acre manufacturing site and moved more than 70,000 cubic yards. The number matters because it is not a demo loop. It is production-shaped evidence, accumulated bucket by bucket.
In February 2026, Bedrock raised $270 million in a Series B co-led by CapitalG and the Valor Atreides AI Fund, reaching a reported $1.75 billion valuation. Total funding passed $350 million. The money is intended to move the company from individual supervised deployments toward initial operator-less excavator work and, eventually, coordinated fleets of different machine types.
Dual bachelor's training at Carnegie Mellon, followed by doctoral work in robot learning and navigation.
Anki Drive appears at Apple's WWDC and makes artificial intelligence move across a living-room floor.
After Anki closes, a 13-person engineering group joins Waymo's autonomous trucking program.
Bedrock introduces its retrofit autonomy system and begins proving it on active construction sites.
A $270 million Series B backs the move toward operator-less equipment and coordinated fleets.
Capacity, not spectacle
Automation stories are often written as contests between people and machines. Sofman's framing is about the work that scarce crews cannot get to. Construction companies turn down projects. Experienced operators spend long shifts repeating cycles at remote sites. Retirement drains knowledge faster than training replaces it. The Bedrock pitch is that autonomy can take repetitive work, keep expensive equipment productive for longer hours, and let skilled people supervise more of the system.
That claim will be tested in dust, rain, scheduling meetings, safety reviews, and the quiet skepticism of operators who know when a machine is behaving badly. Sofman is unusually prepared for that kind of judgment. Anki taught him that a technically impressive robot still has to earn a place in someone's home. Waymo taught him that edge cases do not politely announce themselves. Bedrock brings both lessons to a workplace where reliability is visible in the grade.
“The technology has to earn its place on every job site, one project at a time.”Boris Sofman
There is also a personal loop closing in Sofman's story. His childhood game was about coordinating a civilization's limited resources. His current company is about expanding the resources available to build one. He has said that his own son now plays Civilization, a detail he admits makes him strangely proud. The inheritance is less about a particular game than a habit of looking at a complex system and asking what intelligence might make possible.
The ambition at Bedrock stretches well beyond one excavator. Heavy machines exist in mining, agriculture, forestry, and waste. A common autonomy layer could eventually coordinate different equipment toward a project goal, turning a fleet into something closer to an operating system for physical work. That future remains ahead. For now, the honest unit of progress is smaller: one machine, one task, one site, one pile of earth in the right place.
Sofman's machines have learned to race, gesture, drive, and dig. Each chapter traded a controlled environment for a messier one, and each raised the price of being wrong. The through line is not size or spectacle. It is the patient conversion of intelligence into behavior people can use. At Bedrock, that behavior leaves tracks in the dirt.