AI summary by TwinMind
154. Can Robots Get a Grip? | People I (Mostly) Admire
Freakonomics Radio Network
57m 52s
Early Robotics Inspiration and Career Path
- Goldberg's father ran chrome plating company, built robot for dirty/dangerous work with poisonous chemicals like cyanide
- Father built frame with stepping motors, switches, controller like player piano with rotating drum
- Taught Goldberg binary numbers and LEDs
- Project never fully worked due to father's limited attention span
- Word "robot" appeared in 1920 Czech play, derived from root word for work/forced work
- Goldberg defines robot as programmable machine that moves in physical world, doesn't need to look human
- Studied in Scotland early 1980s, exposed to AI department with pioneers from Alan Turing's work
- Found grasp lab at UPenn led by Professor Ranichi, worked on grasping problems throughout career
Dissertation Research on Grasping Theory
- Focused on picking up objects with parallel jaw gripper without using sensors
- Developed mathematical theorem for constraining polygonal objects to achieve unique final orientation
- Took 2 years to prove completeness theorem for any polygonal part
- Combined theoretical proof with practical gripper invention
- Worked well for orienting objects (part feeding) but small errors and friction violated assumptions
Challenges in Robot Grasping and Manipulation
- Vision problems: 3D depth perception still difficult despite high-resolution 2D cameras
- LIDAR sensors very noisy, can't accurately determine object positions in space
- Autonomous vehicles only need to avoid hitting things, grasping requires precise contact
- Fine motor skills: tactile sensing extremely difficult
- Goldberg's early sensors had false positives (freezing when no contact) and false negatives (continued squeezing)
- Sensor signals vibrating, moving, drifting constantly
- Human hand exquisitely designed with beautiful skin, sensory processing, huge dynamic range
- Mechanical systems inherently have small imprecisions
- Humans compensate elegantly with constant adjustments, fast feedback loops
- Example: pushing pencil on table - outcome undecidable due to microscopic friction variations
Computer Vision Revolution and Neural Networks
- Fei-Fei Li's ImageNet around 2010 created breakthrough in computer vision
- Three key ingredients converged in 2012: data, computation (GPUs), algorithms
- Neural networks started generalizing beyond specific training when given enough data
- GPUs originally developed for games, repurposed for AI computation
The Bitter Lesson and Model-Free Approaches
- Rich Sutton's 2019 essay "The Bitter Lesson" - traditional rule-based approaches consistently outperformed by data-driven methods
- Scientific approach traditionally involved human-designed algorithmic rules
- Neural networks are "black boxes" - don't understand how they work but empirically superior
- Model-free methods just throw data at problems, interpolate and figure out solutions
Data Requirements and the "Data Gap"
- Large language models trained on 1.2 billion hours of data
- Current robot training datasets around 10,000 hours (approximately 1 year)
- Would need 100,000 years to reach language model data levels at current rate
- No equivalent of written language archives for physical manipulation tasks
- Simulation approach has limitations - physics simulators lack real-world friction and imperfections
- Domain randomization (adding appropriate noise) can help bridge sim-to-real gap
Dex-Net Breakthrough (2016)
- Generated large dataset of 3D objects and grasps using Monte Carlo integration
- Added noise/perturbations to make training more realistic
- Looked for robust grasps with highest probability of success under uncertainty
- Never identified objects - only used point cloud patterns to determine gripper placement
- Achieved over 90% success rates, 250 picks per hour (humans ~400 picks/hour at 99.9% success)
- Led to commercialization through Ambii company
- Accumulated 22 years worth of real-world warehouse data during pandemic e-commerce surge
Tesla Autonomous Driving vs Waymo Comparison
- Tesla has ~500x more data than Waymo but worse performance
- Tesla uses end-to-end approach: raw camera images → steering/braking decisions
- Waymo uses multiple sensors (cameras, LIDAR, etc.) plus good old-fashioned engineering
- Demonstrates data alone insufficient - need proper engineering and system design
Concerns About Current Robotics Hype
- Biggest wave of humanoid robot interest in Goldberg's 40+ year career
- Elon Musk claims Optimus robots could generate $1+ trillion revenue
- Goldberg worried about unrealistic expectations leading to AI winter
- Gartner hype cycle: initial excitement → peak → crash → gradual real progress
- Locomotion/walking capabilities impressive but fine motor skills still lacking
- Humanoid form factor more about marketing/perception than functionality
Art and Science Integration
- Goldberg creates art installations, had dozen solo exhibitions at Whitney Museum, Pompidou Center
- Telegarden (1994): internet-controlled robot tending garden, 100,000+ participants
- Motivated by desire to reach broader audience than typical 20-person gallery visits
- Art projects led to NSF grants for telerobotics research
- Recent collaborations: dance performance "Breathless," Getty exhibition wood carvings about scientific timelines
- Two cultures problem (C.P. Snow 1959): scientists and artists mutual disdain/incomprehension
Medical Applications
- Appointment at UC San Francisco radiation oncology department
- Worked on brachytherapy: delivering radioactive seeds to treat cancer
- Analogous problem to grasping - moving objects precisely through space with uncertainties
- Surgical robots are sophisticated puppets with human operators
- Keyhole surgery through small abdominal holes with robot grippers
- Better ergonomics for surgeons enables better concentration and precision
- Focus on augmenting human capabilities rather than replacement
Future of Human-Robot Collaboration
- 100% believes in complementarity over substitution
- Technology history shows augmentation rather than replacement
- ChatGPT example: helps humans think better, doesn't replace thinking
- Shortage of human workers, robots won't put people out of work
- Not afraid of singularity - believes humans will remain in control
- Some cases of AI running amok but not existential threat level
