AI summary by TwinMind
Rodney Brooks: Robotics | Lex Fridman Podcast #217
Lex Fridman
2h 24m 52s
Introduction and Background
- Conversation with Rodney Brooks, one of greatest roboticists in history
- Led MIT AI Lab, co-founded iRobot (most successful robotics company), Rethink Robotics (collaborative robots Baxter/Sawyer), Robust.ai (teaching robots common sense)
- Host notes 20+ year journey in robotics, agrees/disagrees with Brooks' AI critiques
Most Beautiful Robot - Domo
- Domo created by grad student Aaron Edsinger, now in Daniela Rus's office
- Beautiful upper torso humanoid, two arms, three-fingered hands, face with eyeballs
- Series elastic actuators, cable-driven, motors inside, mechanically gorgeous
- Aaron didn't give budget ahead of time, spent heavily with Jeff Weber (mechanical engineer)
- 512 bytes RAM, exquisite detail, anodized parts in different colors
- Visual attention mechanism - looked at people, tracked faces
Robot Design Philosophy
- Robot appearance makes promise for interaction capabilities
- Encourages students not to over-promise with design
- "If robot looks like Albert Einstein, should be as smart as Albert Einstein"
- Domo only had eyeballs because that's all it could do - look and pay attention
- Baxter/Sawyer had graphic eyes on screen showing intent before arm movement
Early Robotics Journey
- Born 1954, grew up Adelaide, Australia
- Two books from 1961: "How and Why Wonder Books" on electricity and giant brains/robots
- Age 8-9: learned circuits, binary system, tried building robots
- Built learning system with ice cube tray - chemical-based neural network using copper bridges
- Later built transistor-based gates and flip-flops
- First "thinking" robot was ice cube tray that learned to add
Philosophy on Machine Intelligence
- Machines can think because humans are machines and think
- Question is whether we're capable of building such machines
- Writing 480-page book "Not Even Wrong" with three thrusts
Book: "Not Even Wrong" - History of Computation
- Traces computation history from 1614 Napier/Kepler manuscripts through Babbage/Lovelace to Turing 1936
- Turing's 1936 paper aimed to disprove Hilbert's hypothesis, not invent computation
- Modeled after human "computers" who followed mechanical processes with paper/pencil
- Turing defined computation based on intuition, not rigorous proof
- From 1936-1975, mathematicians struggled with whether this was really computation
- Minsky, Knuth, others acknowledged uncertainty about computation definition
Four-Quadrant Matrix of Related Fields
- Science row: Neuroscience, Abiogenesis (how non-living becomes living)
- Engineering row: Artificial Intelligence, Artificial Life
- All four disciplines emerged 1945-1965 with common players
- Warren McCulloch central figure at MIT Research Lab for Electronics
- McCulloch-Pitts 1943 paper on neural calculus inspired by Bertrand Russell
- Three disciplines adopted computation as primary metaphor
Computation vs. Other Processes
- Computation fits hunter-gatherer thinking: places, containers, stuff in places
- Example: drum vibration finds stationary points without computing
- Ada Lovelace's first program computed Bernoulli numbers (needed for drum stationary points)
- Questions whether computation is right metaphor for intelligence/life
Intelligence and Embodiment
- Every intelligence we know perceives world and acts in it
- Perceive objects in sophisticated ways (color constancy, context-dependent perception)
- Deep learning systems miss this registration between perceiver and world
- Symbol grounding problem not solved by image labeling - different from true grounding
- Intelligence likely requires interaction with other humans, external tools
Moravec's Paradox
- Reasoning appears easy but perception/mobility are hard
- Reflects bias of intelligent humans who were good at reasoning tasks
- Evolution spent 600M years on perception/movement, recent time on language/reasoning
- Chess/reasoning built on top of earlier perceptual foundations
Manipulation and Learning
- Story of 16-month-old grandson opening window with handle - two different hands, different actions
- Demonstrates sophisticated understanding without reinforcement learning trial-and-error
- Kids don't randomly try everything - apply pre-filters that dramatically reduce search space
- Possible internal simulation/prediction, but definitely not pure random search
Deep Learning and Neural Networks
- Shocked by ImageNet performance - didn't expect that level of computer vision capability
- Alpha Zero impressive but techniques existed in Mickey's 1961 paper
- Questions whether this represents fundamental breakthrough vs. better implementation
- Arthur Samuel 1960s checkers program beat world champion using similar self-play
- Samuel had unique advantage: access to IBM's entire computer farm for testing
Self-Driving Cars - Technical Progress
- Impressed by Mercedes E450 capabilities: lane keeping, 360-degree view, blind spot detection
- Surprised by Mobile Eye's computer vision performance initially
- More surprised when Tesla caught up after dropping Mobile Eye
- Andre Karpathy's data engine approach with edge case discovery impressive
- Continuous surprise at vision-based methods' capabilities
Self-Driving Cars - Deployment Challenges
- 35,000 US traffic deaths/year, but public won't accept 10+ autonomous vehicle deaths
- Need infrastructure changes like trains (dedicated lanes, controlled environments)
- Expected dedicated highway lanes by now, hasn't happened
- Cameras lack human eye's 11 orders of magnitude dynamic range
- Long tail of edge cases very long, acceptance standards much higher than human drivers
Train Systems and Infrastructure
- 15-16 fully autonomous train systems in US, mostly airports
- Honolulu system expected 2017, delayed
- Autonomous trains require infrastructure: double doors, protected tracks
- Expects similar infrastructure changes needed for autonomous vehicles
Predictions and Timelines
- Maintains prediction blog since January 2018, reviews annually through 2050
- Cambridge Massachusetts full deployment: 2035 (may be too optimistic)
- San Francisco grid streets north of Market: still few years away with designated pickup/dropoff
- South of Market much longer without road re-engineering
- Edge case example: Cruise vehicle blocking crosswalk for extended time
iRobot Success Story
- 14 failed business models before 2002 success
- Roomba and military robots successful same year
- 30+ million household cleaning robots sold
- Proud of Fukushima nuclear plant robots (March 2011)
- 6,500 military robots deployed Iraq/Afghanistan gave harsh conditions experience
- Japanese robots got hype but couldn't perform when needed
Roomba Development Details
- Aimed for $200 consumer price vs. $2,000 Electrolux Trilobite
- Extremely tight budget required innovative sourcing
- Traveled to Taiwan/Hong Kong finding stripped-down processors
- Found Winbond chip with 512 bytes RAM within budget
- Built 70,000 units first year (3x board authorization), sold all by Christmas
Rethink Robotics - Vision and Failure
- Goal: robots out of cages, safe human-robot collaboration
- Humans could re-task without coding
- Original vision: $3,000 plastic robots for new market segment
- Forced into metal gearboxes by engineering team complaints
- Price jumped to $25,000 Baxter, then $35,000 Sawyer
- Now competing with traditional industrial robot expectations
- Failed due to CFIUS blocking Chinese acquisitions, asset sale at 1/30th price
Robotics Company Challenges
- Founder expectations of capabilities often off-base
- Mispricing relative to customer value perception
- Customer adoption difficulty for new technology
- Expects demographic inversion will drive home robotics demand
- Significant capital now available (saw $80M Series B for Covariant)
MIT AI Lab Memories
- Regrets not asking more questions of AI pioneers who have passed
- Would ask Licklider about human-computer collaboration vision, successes/failures
- Would ask Minsky why he attacked neural networks in 1968 Perceptrons book after doing neural network PhD thesis
- Access to future technology: video terminals, digital sound, graphics printing, Lisp machines
- Email since 1977, early internet connectivity
Donald Knuth Story
- Knuth worked overnight at AI lab writing TeX
- Brooks and Robert Paul fixed mainframe disk drive failure
- Nervous with tall Knuth standing over them asking when it would be fixed
- Put chip in backwards, smoke came out, eventually fixed it
Advice for Young People
- No single path to success
- Important not to follow everyone else
- Real impact requires readiness to fail repeatedly
- Make unsafe decisions over long time periods
- No way to make all safe decisions and truly contribute
Mortality and Legacy
- COVID made him work harder on book initially
- Fears losing mental acuity, wants to finish "Not Even Wrong"
- Hopes one person reads book and changes their thinking
- That person might say on future podcast "I read that book and it changed everything"
Meaning of Life and Universe
- Atheistic view: existence is random
- Interested in how order emerges from disorder
- Impossible to believe we're alone in universe
- Most likely outcome: graveyard of intelligent civilizations everywhere
- Most people forgotten within generation or two
- Despite absurdity and meaninglessness, life is fun
