AI summary by TwinMind

Robert Playter: Boston Dynamics CEO on Humanoid and Legged Robotics | Lex Fridman Podcast #374

Lex Fridman

2h 27m 58s

Early Career and Natural Walking Development

  • Goal was natural looking gait for humanoid robots - surprisingly hard to achieve
  • Built early machine called "Petman prototype" before Atlas robot
  • Nice looking gait with heel strike first, rolling onto toe (not flat foot landing)
  • Difficult to get robot to fully extend leg during walking
  • Natural walking took 15 years (10-15 years) to achieve
  • Petman prototype: ~2008, good walking on Atlas: ~2022
  • Team developed new techniques for walking control algorithm on newer Atlas generation
  • Natural motion achieved as byproduct of different development process

Background and MIT Leg Lab Experience

  • Robert Playter: CEO Boston Dynamics, PhD Aeronautical Engineering MIT 1994
  • MIT Leg Lab - wrote PhD thesis on robot gymnastics
  • Programmed bipedal robot to do world's first 3D robotic somersault
  • First fell in love with robotics visiting MIT, saw Mark Raibert's robots doing somersaults
  • Immediate connection due to personal gymnastics background
  • Core lessons from Mark Raibert:
    • Have courage of convictions, pursue what's interesting
    • Don't worry about making money initially
    • Try to simplify problems to core essence
    • It's okay to break stuff - test, fix, repeat approach

Philosophy of Robot Movement and Design

  • Fascination with movement and letting body do what it wants naturally
  • In gymnastics: putting body in position where physics/inertia push in right direction
  • Building machines to take advantage of natural physics
  • Spring-mass oscillating systems with natural frequency
  • Coordination between natural physics and intentional control creates "beauty"
  • Efficiency through working with physics rather than overpowering them
  • Physics can provide stabilizing mechanisms without computer control

Company History and Evolution

  • Boston Dynamics started 1992
  • First 10-11 years: simulation company, not robotics
  • 2003: DARPA contract to build robot dog (Big Dog)
  • Transition from simulation to building actual robots in ~1 year
  • Big Dog: quadruped with go-kart engines, hydraulic actuators, very loud
  • Proved legged locomotion could work in self-contained robot
  • Led to series of robots: LS3 (800 lb robot, 400 lb payload), Cheetah (speed exploration)
  • Developed design principles for scaling robots by size, payload, terrain

Spot Development and Commercialization

  • Born ~2012 after 10 years of DARPA quadruped development
  • Transition needed: smaller, electric (not gas), not hydraulic
  • Google acquisition - Larry Page wanted consumer product ("toothbrush test")
  • Philosophy difference: Boston Dynamics wanted high-quality machines first
  • Focus on industrial sector - can afford more expensive machines
  • Current price point: ~$100,000+ acceptable for industrial productivity
  • Path to consumer products through industrial applications and mass manufacturing

Atlas Humanoid Capabilities

  • Recent video showing manipulation: throwing heavy bags, using tools
  • Explicit modeling of objects being picked up (weight distribution, center of mass)
  • Robot anticipates center of mass shift when picking up objects
  • Can create new behaviors in days (previously took 6 months)
  • Box jump capability: ~1 meter high
  • Next generation Atlas will have stronger actuators than humans can achieve

Jumping and Acrobatic Control

  • Early backflips required manual iteration and trajectory adjustment
  • Now using model predictive control - robot thinks 1-2 seconds ahead
  • Can adjust on the fly during airborne maneuvers
  • Front somersault with half twist - robot adapts to different landing configurations
  • Real-time projection and option exploration during flight

Manufacturing and Cost Challenges

  • Transition from R&D to commercial company
  • Manufacturing facility in Waltham, ~1 mile from headquarters
  • Learning: casting parts vs milling from billet aluminum, plastic molded parts
  • Challenge: still iterating design while scaling manufacturing
  • Supply chain complications when changing designs after bulk ordering
  • Engineers adapting to work on quality, reliability, cost problems

Current Fleet and Reliability Testing

  • ~1100 Spot robots deployed with customers
  • Need 1000-1500 robots/year for profitability
  • Dozen robots walking around Boston Dynamics offices continuously
  • Weekly metrics: 1500-2000 kilometers walking, 1000+ hours operation
  • Autonomous missions: get up, walk building, collect data, return to dock
  • Essential for finding rare failures and improving reliability

Stretch Warehouse Robot

  • Mobile robot arm on pallet-sized base
  • 16-hour battery life (two shifts)
  • 50-pound payload capacity
  • First deliveries January 2023: DHL, Maersk, Gap
  • 7-8 additional customers committed to 10-20 robots each
  • Addresses trillion cardboard boxes shipped annually
  • Customers see 2-year return on investment

Handle Robot (Discontinued)

  • Balancing robot on wheels with legs, arm, and tail
  • All components moving simultaneously for balance coordination
  • Beautiful, graceful movement but too slow for logistics
  • Business decision to abandon for more efficient Stretch design
  • Demonstrates what's possible with momentum control

Competition and Tesla

  • Elon Musk/Tesla humanoid robot announcement brought validation
  • Proliferation of robotics companies claiming Boston Dynamics alumni
  • Competitive energy driving Atlas improvements
  • Recent Atlas video showing manipulation + parkour capabilities
  • Focus on heavy object manipulation (50 lbs for 150 lb robot)
  • Next phase: more dexterous hands, two-armed coordination

Technology and Learning

  • Reinforcement learning used for locomotion, manipulation, perception
  • Physics-based simulation tools developed since MIT days
  • Same code runs in simulation and hardware
  • 40 developers sharing 4 Atlas robots
  • API allowing partners to add vision algorithms
  • Ecosystem approach for specialized capabilities

AI and Large Language Models

  • Physical reality provides verification mechanism for robotics AI
  • Different from language models where truth is harder to verify
  • Partner demonstrations of language tools integrated with Spot
  • Excitement about conversational capabilities with robots
  • Less concerned about AI risks in robotics vs pure software applications

Safety and Weaponization

  • Authored letter with leading robotics companies pledging not to weaponize robots
  • Co-signers: Unitree (China), Agility (US), ANYmal (Europe), others
  • Concern about commercially available robots being modified with weapons
  • Importance of trust for industry growth
  • Policy director engaging with regulators and lawmakers

Future Vision and Social Robotics

  • Path to consumer robots through industrial applications first
  • Performance and utility must come before social features
  • Potential for robots as companions/pets in long term
  • Physical gesture communication enhancing human-robot interaction
  • Complementary work relationship, not replacement of all human labor
  • Robots handling dangerous/undesirable tasks while humans do creative work
  • Optimism about robots addressing loneliness through companionship

Leadership and Hiring Philosophy

  • 30 years at one company with constantly evolving role
  • Key traits: courage, adaptability, willingness to tackle any business need
  • Interview process: ask about candidate's interests and expertise
  • Probing questions to verify actual work done and passion
  • Presentation requirement for all roles to assess communication
  • Focus on finding genuine curiosity and care for the work

Boston Dynamics AI Institute

  • Separate organization led by Mark Raibert
  • Focus on research without commercialization constraints
  • Longer time horizon than Boston Dynamics (beyond 5 years)
  • Hyundai investment in pushing technological frontiers
  • Collaboration expected between commercial and research arms

TwinMind

Get summaries like this for your own meetings

TwinMind takes notes in the background during meetings, lectures, and calls, then writes the summary for you.

Free plan to get started

No bot joins your calls

Phone, desktop & Chrome apps

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome