Sofman, Boris

Senior Director of Engineering and Head of Trucking at Waymo; co-founder and former CEO of Anki, creator of the social robot Cozmo

Boris Sofman: Waymo, Cozmo, Self-Driving Cars, and the Future of Robotics

Lex Fridman Podcast · Nov 16, 2021

• Boris Sofman
• Anki and Cozmo social robot
• human-robot interaction design
• Waymo Via autonomous trucking
• Waymo vs. Tesla self-driving approaches
• lidar vs. vision sensor fusion
• autonomous vehicle safety and evaluation
• trucking industry automation and labor impact
• humanoid robotics skepticism
• startup failure and emotional toll

Core ideas

Sofman explains that Anki's founding thesis — developed with two co-founding PhD students at Carnegie Mellon's Robotics Institute — was that consumer robotics was almost entirely unexplored territory outside research and industrial/military applications, and that the falling cost of mobile-phone-era components (cameras, microcontrollers, sensors) made it newly possible to push a robot's complexity into cheap hardware plus rich software, much like "bringing a video game to life" in the physical world. Anki's first product, the racing game Drive, proved that experiences which feel trivial virtually have outsized emotional impact when physically real; this insight directly motivated Cozmo, an intentionally non-humanoid character (only four degrees of freedom: head lift, arm lift, two tank treads, plus sound and a low-resolution screen) modeled explicitly on Pixar's WALL-E, built by a team that included former Pixar animators. Sofman argues the deliberate avoidance of human form was strategic, not just cost-driven: humanoid form imposes expectations that AI's strengths must match human strengths symmetrically, whereas minimalist, non-human characters (citing WALL-E, R2-D2, and simple stick-figure animation) can communicate an equally full emotional range while playing to AI's actual strengths and sidestepping its weaknesses — and that mistakes made by a character with personality read as endearing rather than as robotic failure, a dynamic he considers one of the deepest lessons for human-robot interaction generally.

He describes Cozmo's design as balancing three elements — mechanical/industrial design, AI, and character — arguing most robotics companies over-invest in mechanical sophistication and under-invest in character, which paradoxically raises costs and expectations without guaranteeing an emotionally successful product. Key design choices included building a parameterized, randomized library of extreme emotional responses (rather than fixed animations) so behavior felt spontaneous within believable bounds, discovering that simple upticks in eye-contact frequency measurably increased engagement (a 40% play-time increase from roughly doubling eye-contact rate), and deliberately omitting speech entirely — arguing that giving a character a voice immediately narrows its appeal to a specific demographic and that current speech technology isn't good enough to avoid breaking the illusion of aliveness, whereas non-verbal characters (again citing WALL-E, R2-D2) generalize across ages and cultures. He notes an unexpected finding that a physically embodied, characterful device dramatically lowered users' privacy concerns (contrasted with the skepticism anticipated, and indeed encountered by other tech products, in privacy-sensitive markets like Germany and France) because the character's visible attention and "intent" made its sensing feel natural rather than intrusive — a dynamic he compares to how humans read each other's eye contact and body language.

Sofman discusses Anki's 2019 shutdown candidly and emotionally: despite strong commercial results (millions of units sold, ~$100 million annual revenue, category leadership), the business was brutally seasonal (85% of volume in Q4) creating extreme cash-flow swings that became harder to manage precisely as the company scaled, compounded by a broader collapse in hardware-company valuations (citing Fitbit and GoPro) and a last-minute financing round falling through. He describes it as one of the most emotionally difficult periods of his life — repeated near-death cash crunches, ultimately breaking down crying when announcing the wind-down — while noting that very few employees left during the crisis and that the shutdown night ended with the team celebrating and telling stories until early morning, reflecting genuine pride in what they'd built; he still has an unopened box of fan mail from Cozmo customers he hasn't been able to bring himself to read. He connects this to a broader observation that Anki's AI/robotics team (with directly relevant self-driving-adjacent backgrounds) largely moved together to Waymo afterward.

Turning to his current work leading Waymo's trucking effort (Waymo Via), Sofman explains Waymo's "driver" concept as a platform-agnostic autonomy system (analogous to a human driver's license not being tied to a specific vehicle) that spans both passenger transportation (Waymo One) and freight (Waymo Via). He describes the initial trucking strategy as using "transfer hubs" at metro-area entry points to intentionally simplify the problem — automating hub-to-hub freeway routes first while handling the "last five miles" manually, rather than solving the much harder general surface-street problem immediately — comparing this staged rollout to how Waymo achieved passenger-vehicle autonomy first in a single city (Phoenix) before expanding. He details truck-specific engineering challenges distinct from passenger cars: dramatically different braking/rollover dynamics for an 80,000-pound loaded trailer, jackknife risk, load-dependent handling that must be fed into planning algorithms, and a sensor layout using two outboard pods (near where side mirrors would be) rather than a single roof-mounted pod, to avoid trailer occlusion — while stressing that a large majority of the underlying perception, compute, and behavior-prediction stack (he cites Waymo's models reusing ~85% car-derived data even for truck behavior prediction) transfers directly from the car program, giving Waymo significant leverage across both product lines.

On evaluation and safety, Sofman argues the hardest problem in autonomous driving is not autonomy itself but evaluation — since safety-critical failure events are extremely rare (a serious truck freeway accident occurs roughly once per 1.3 million miles, a severe injury once per 28 million), and Waymo simulates roughly 1,000 miles for every mile physically driven to enable much faster iteration, regression testing, and stress-testing of edge cases without incurring real-world risk. Invited to compare Waymo's sensor-rich (lidar, camera, radar) L4-focused approach against Tesla's vision-only, mass-fleet-data strategy, he argues Waymo's advantage is architectural focus on full autonomy from the start (avoiding a costly hardware/software leap later from an L2 assistance system), while granting Tesla's real advantage is genuinely massive real-world data collection — with the central open question being whether pure computer vision can, and how soon it will, match the reliability lidar provides for rare edge cases, given that adding perception risk compounds an already extremely hard behavioral/planning problem. Fridman pushes back somewhat, suggesting society may eventually need to accept incrementally higher (though still human-superior) risk in exchange for autonomy's benefits the way it tolerates human driving risk-taking, and Sofman agrees this is plausible in the long run but stresses the current ethical bar is necessarily much higher than for a personal risk decision, since autonomy providers make risk choices affecting other road users, not just themselves.

On trucking's economic and labor impact, Sofman argues trucking is unusually positioned to benefit from automation with limited net-negative labor disruption: there's already an 80,000+ driver shortage projected to grow into the hundreds of thousands, current pay structures often only compensate drivers for active driving time (penalizing logistics inefficiency), and the hardest, most dangerous long-haul routes autonomy would take over first are also the least desirable jobs (300+ days away from family, high obesity and fatigue-related risk) — while the more desirable, human-relationship-dependent short-haul and last-mile routes are likely to remain human-driven far longer. He cites an Amazon Kiva-robot warehouse example (where automating item retrieval increased overall throughput enough that final-order-fulfillment jobs held roughly steady even as search/retrieval was automated) as a model for how automation can be a non-zero-sum transition rather than pure job elimination, while acknowledging this doesn't eliminate real short-term human disruption and requires intentional retraining and transition support.

Discussing broader home/humanoid robotics, Sofman is skeptical Amazon's Astro will succeed at its ~$1,500 price point absent a clearer "killer" value proposition, and is generally skeptical of humanoid robot forms (including Tesla's humanoid robot project) for most applications, arguing wheels and non-human joint configurations are usually more efficient, and that humanoid form is often "a technology in search of an application" — though he grants a countervailing argument (raised by Fridman) that humanoid form could be compelling specifically for social integration/companionship rather than factory efficiency, similar to Cozmo's appeal. He closes with career advice emphasizing finding the overlap between personal passion, a growing market opportunity, and a durable skill set (citing machine learning as a versatile foundation applicable across many careers), and reflects on having sometimes worked to unhealthy extremes, now consciously trying to balance intense work with long-standing friendships and family.

Excerpts