← Back to list
Lex Fridman PodcastPodcast16 Nov 2021Source: lexfridman.comHost: Lex Fridman

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

In plain words

This interview reveals the hard truths of robot startups and self-driving cars. Boris (a Waymo executive) says picking the right market matters more than great tech—his old robot company Anki failed because the toy market is tiny and seasonal, while trucking is huge. He favors Waymo’s multi-sensor approach (LiDAR + cameras) over Tesla’s pure vision. Key mentions: Waymo (runs 1,000 simulated miles for every real mile to test safety), Tesla (collects tons of real data from customers but struggles with unusual conditions), and Anki (its robots had ‘character’ that made people forgive mistakes, but the market was too small).

AI SummaryAI-generated · may contain errors · verify against the original

This research summary draws on an interview with Boris Sofman, Senior Director of Engineering and Head of Trucking at Waymo, discussing autonomous trucks, robotics technology, and his entrepreneurial experience at Anki. Core insights: Boris co-founded the robotics company Anki, which developed the s

~8 min full read · 10 sections
Deep Analysis

[SKIP]

Core Insight: The Investment Logic and Technical Philosophy from Cosmo to Waymo

1. The Immutable Law of "Market Fit" for Robotics Startups: Great Technology Is No Match for the Right Sector

When summarizing lessons from Anki, Boris emphasized the decisive role of market timing and industry dynamics. He pointed out that even with a top-tier team and technology, if the industry's "market capacity" and "capital tolerance" are insufficient, success is unlikely. For example:

  • Anki's Dilemma: The consumer entertainment robotics market is highly seasonal (85% of sales in Q4) and requires convincing both children and parents simultaneously, leading to high marketing costs and immense cash flow pressure. Even with excellent product reputation, the "laws of physics" cannot be changed — namely, the conflict between low valuation multiples for hardware companies (1x revenue) and enormous working capital needs.
  • Waymo's Opportunity: The autonomous trucking market is hundreds of billions of dollars in size and extremely capital-intensive, requiring massive early investment but also offering enormous potential returns. This "winner-takes-all" track makes huge investments logically reasonable — an "abnormal" but justified capital allocation.

Core Conclusion: When starting a business, choosing a sector with a large market size, rapid growth, and capital willing to "tolerate" long-term losses is more important than pursuing technical perfection alone.

2. The "Long-Tail" Safety Assessment of Autonomous Trucks: Harder Than the Technology Itself

Boris delved into the assessment problem as one of the biggest challenges facing autonomous driving — essentially a "meta-problem." He provided specific data:

  • Human Baseline: The probability of a truck driver being involved in a serious accident on a highway is approximately 1 in 1.3 million miles, while a fatal accident is as high as 1 in 28 million miles. This means that even if a system appears to perform perfectly, actual road testing cannot confirm its safety level because rare events occur too infrequently.
  • Waymo's Approach:
  • Simulation: For every actual mile driven, there are 1,000 miles of simulation testing. Simulation can reproduce long-tail scenarios and accelerate the exploration of safety boundaries.
  • Structured Testing: Forcing failures (e.g., sensor failure, tire blowout) and verifying the system's response in a controlled environment.
  • Correlation Metrics: Finding intermediate metrics that are highly correlated with ultimate safety but easier to measure (e.g., proximity to human driving), enabling rapid iteration.

Breakthrough Point: Boris noted that the "Turing Test" of assessment is not driving a section of road, but whether a system can pass a carefully designed set of "IQ tests" containing extreme scenarios. This echoes François Chollet's view on machine intelligence testing — the real challenge is generalization to unseen complex situations.

3. Waymo vs Tesla: A Comparison of Technical Routes and Data Strategies

In his comparison, Boris provided very specific arguments. We can present them in a table:

Dimension Waymo Tesla
Core Goal Start from L4, customized hardware and software, pursue system-level safety proof Evolve from L2+, focus on massive data collection and rapid iteration
Sensor Strategy Multi-sensor fusion (LiDAR + camera + radar), early fusion Vision-only (remove radar), rely on cameras
Data Advantage Millions of miles of real road testing + 20+ billion miles of simulation Millions of vehicles already sold, generating massive amounts of real-world road data daily (and paid for by users)
Risk Point High sensor cost, high system complexity, but high redundancy Long-tail challenges of visual perception (e.g., strong light, fog, abnormal objects) could become a safety bottleneck
Iteration Speed Limited by L4 safety validation, slower release pace Can be quickly updated via OTA, but may lead to safety risks if overly aggressive

Boris believes that a pure vision system is theoretically feasible (humans are an example), but solving the computer vision "long-tail" problem under L4 safety standards would make an already difficult autonomous driving problem even more challenging. Waymo's sensor fusion (especially self-developed LiDAR) provides more reliable physical consistency, helping to achieve a better safety boundary statistically.

4. The "Hardware-Software Balance" in Robotics Startups: Character is the Core Moat

Boris repeatedly emphasized that in Anki's experience, "character" is the key to building trust between robotic products and humans. Specific manifestations include:

  • Mistakes Become Endearing: When a robot makes a mistake, if it can show an emotional response indicating it "realizes the mistake" (e.g., frustration, confirmation), it actually enhances empathy rather than causing disgust.
  • Privacy Concerns Disappear: Because Cosmo has "character," users naturally accept that it needs a camera and microphone — "how else would it come to life?" This is far more effective than any cold privacy statement.
  • Implications for the Home: This points the way for future home robots and voice assistants (e.g., Amazon Astro): mistakes by robots without character are amplified; mistakes by robots with character are tolerated.

Boris believes that most current robotics companies (including tech giants) over-invest in hardware and AI but severely underestimate the importance of "character" design. This may be a key missing piece for the robotics industry's explosion in the next decade.

5. A Sober Judgment on Humanoid Robots: Technically Feasible, but No Clear Application

Regarding Tesla Bot, Boris gave a clear view:

  • Current Stage: Humanoid robots are research-oriented and have yet to find a killer application that matches their cost.
  • Comparison with Cosmo: The human form brings extremely high expectations (e.g., language, movement, intelligence), and any deviation triggers the "uncanny valley" effect. Non-humanoid robots like Cosmo, on the other hand, can more freely leverage their strengths and avoid weaknesses.
  • Future Possibility: If autonomous driving technology (e.g., Waymo Driver) matures, its perception, planning, and control capabilities could be transferred to more general robotic platforms, but not necessarily humanoid. For example, cargo handling could use a wheeled chassis + robotic arm, which is more efficient.

6. An Optimistic View on AI and Employment: Efficiency Gains Create More Jobs, Not Fewer

Using Amazon's Kiva robot warehouse as an example, Boris argued that automation has not led to job losses, but rather:

  • Efficiency Gains: Warehouse throughput has more than doubled, but the number of positions has remained unchanged (work content shifted from "looking for items" to "picking items").
  • Quality Improvement: Among truck drivers, the most exhausting long-haul routes (300 days/year away from home) will be automated, while short-haul, flexible local routes (enabling daily returns home) will still require humans, potentially with higher wages.
  • System Growth: When logistics efficiency improves, the overall market size expands, creating more logistics-related jobs.

Key Point: Automation is not a zero-sum game, but "enlarging the pie." However, proactive government and corporate efforts in retraining and social support are needed to alleviate short-term pain.

7. Personal Advice: Deepen Expertise in Growing Fields While Balancing Life

Boris gave two core pieces of advice:

1. Choose a Growing Field: Preferably one that overlaps with your passion but is also in a rapidly growing industry (e.g., machine learning, AI applications). This way, even if your specific job changes, your skill set remains valuable and opportunities are more abundant.

2. Balance Work and Life: He admitted to overworking in the past but now values friendship, family, and long-term relationships. Don't wait until retirement to build these relationships, because life cannot be "replayed."

Data Comparison Table: Waymo Simulation vs Real-World Testing

Metric Value
Real-world testing miles Over 20 million miles
Simulation miles Over 20 billion miles (i.e., 1,000 miles of simulation for every real mile)
Interval between human truck driver serious accidents Approximately 1.3 million miles
Interval between human truck driver fatal accidents Approximately 28 million miles

Summary

Boris Sofman's interview reveals the profound challenges of robotics entrepreneurship and autonomous driving from technical, commercial, and philosophical perspectives. The core idea is: engineering success does not equal commercial success. Choosing the right market, timing, and business model is more important than solving technical problems. Meanwhile, in robot-human interaction, "character" is the key to building trust, and this is a critical variable for the future explosion of the robotics industry.