AI summary by TwinMind

The Humanoid Robot Dream

Asianometry

16m 18s

OpenAI Investment in 1X Technologies

  • OpenAI led $23M investment in Norwegian humanoid robotics company 1X Technologies
  • Company formerly called Halodi, unrelated to Wyoming cable company with same name
  • Investment aligns with OpenAI's early product goal: household robot capable of setting/clearing tables
  • Goal somewhat forgotten due to ChatGPT focus

Humanoid Robot Definition and Requirements

  • No widely accepted definition exists
  • Key characteristics from design purpose:
    • Move around and perform broad spectrum of tasks
    • Physically touch people without harm
    • Manipulate human-designed tools
    • Autonomously operate in human environments
    • Communicate at advanced level
  • Design requirements dictate human-like form:
    • Torso, head, two legs, two arms
    • Hands with multiple fingers
    • Two legs preferred over four for energy, humanity, cost reasons

Industrial vs Humanoid Robots

  • Industrial robots: nailed to ground, specific tasks, big/heavy/fast
  • Not meant for regular human interaction
  • Humanoid robots: operate in tight intimate spaces with humans
  • Requires special engineering considerations

Historical Development

Early Theoretical Foundation (1968)

  • Zero Moment Point (ZMP) stability concept introduced by Miomar Vukobratovich
  • Covers dynamics and control of legged robots with powered controllable joints
  • Critical for balance: robot supports itself on single foot 80% of walking time
  • ZMP represents ground point where robot's footing can be stable

Waseda University Research (1973-1984)

  • Dr. Ichiro Kato's team started humanoid robot research based on prosthesis work
  • 1973: WABOT-1 debuted - walked on two legs, held things, communicated in Japanese
    • Mental faculties judged equivalent to 1.5-year-old
  • 1984: WL-10RD - first dynamically balanced bipedal robot
    • First physical demonstration of ZMP concept feasibility
  • WABOT-2: could sight-read music, play electric organ
    • Guided by 80 microprocessors, mechanically produced using wires

Honda's Breakthrough (1986-1996)

  • 1986: Honda started humanoid robot research program
  • Goal: sophisticated robot as work partner/servant
  • First robot E0 (electronics): walked very slowly (5-30 seconds between steps), straight line only
  • 1996: P2 introduced - fully self-contained, walked on flat surfaces and stairs
  • P2 breakthrough: proved bipedal humanoid robots were possible (many were skeptical)
  • Evolved into government-sponsored humanoid robotics project for domestic helper robots

Current Technical Limitations

Walking Challenges

  • ZMP-based robots have recognizable characteristics:
    • Heavy flat-footed walk
    • One foot always in full ground contact when other is airborne
  • Human body: 300 degrees of freedom, uses up to 20 while walking
  • Robots: simplified with frozen/rigidly controlled joints → unnatural movement
  • Common result: bent knee, frozen torso walking
    • Keeps core stable, easier to calculate
    • Extremely tiring for robot and human
  • Lack of "toe off" behavior (trailing big toe pushes off ground)
    • Important energy-saving feature missing in most robots

Obstacle Navigation

  • Previous walking robots couldn't handle unexpected obstacles >6% of leg length
  • Shorter than average curb height
  • Issues calculating new coordinates for unexpected obstacles on-the-fly
  • Recent progress using AI for approximate results without full model calculation

Recent Advances (2018)

  • French team: robot climbing normal different-sized steps without tethers
  • American team/Boston Dynamics: taught Atlas straight-leg walking
  • Straight-leg walking and toe-off behavior emerged naturally after biasing controller
  • Bird-inspired robots (e.g., Cassie by Agility Robotics):
    • Run 100 meters in under 25 seconds
    • Better navigate unstable terrain
    • Act/run like birds (dinosaurs) - affects human interaction

Manipulation Challenges

Moravec's Paradox

  • Things humans find instinctively easy (clearing table) are very difficult for robots
  • Example: picking up dirty plate requires:
    • Optimal trajectory approach
    • Fingers slide along plate edge, roll under
    • Push hand/fingers forward for grip
    • Manipulate, tip up, lift
  • Requires exceptional mental calibration and soft touch

Soft Actuation Technology

  • Being gentle requires: precise feedback, acute physical awareness, controlled stiffness
  • Machines good at: strong, rigid, powerful
  • Not good at: soft, huggable
  • Soft actuation remains developing field requiring expertise in:
    • Materials, structural design, fabrication
  • Computationally intense: soft material (often silicone) has infinite dimensions
  • Researchers assume rigid rather than soft body conditions → errors

Sensor Limitations

  • Need lots of sensing and feedback
  • Sensors must be cheap and easily manufactured
  • Experiments used: embedded liquid metal, rare earth magnets, optical fiber
  • Challenge: balancing price and sensitivity

Power and Manufacturing Issues

  • Most common soft actuators use pneumatics
  • Require power-hungry, loud, high-vibration support systems (pumps)
  • Industrial arms can pick up plates well (gentleness aside)
  • Humanoid robots need to perform complete sequences under uncertain conditions

AI and Robotics Integration

Adaptive Skill Coordination (Facebook Research)

  • Chains together lower-level individual skills
  • Navigate toward bigger long-term goals
  • Allows robot to back up and correct when things go wrong
  • Trained four-legged Boston Dynamics robot for pick-and-place tasks between apartments

Simulation Training

  • Video game technologies produce simulated worlds for training
  • Huge boon for field - create simulated worlds without actual robots
  • RoboCup 3D Soccer Simulation: uses Nao humanoid robot as agent
  • Generates data for low-level behaviors: walking, standing, turning, kicking

Large Language Models (LLMs)

  • Application not entirely obvious - LLMs trained on text, no physical bodies
  • Not aware of physical context → strange answers (suggesting vacuum for spilled Coke when none nearby)
  • Need to provide context through prompt engineering
  • Google research: trained language model with low-level skills (body + brain)
  • Microsoft suggestion: use LLM to translate ordinary communications to machine-understandable code/API
  • ChatGPT still provides incorrect answers - human must remain in loop

Economic and Technical Challenges

Complexity Comparison

  • Humanoid robots more challenging than self-driving cars:
    • Way more degrees of freedom
    • More intimate human interaction
    • Unknown physical form (vs known car form)
  • Economics complicated - price and manufacturability critical but barely touched
  • Critical points that matter significantly

Investment Interest

  • Funding events for companies: Aptronic, Agility, 1X, Sanctuary
  • Some machines starting to enter warehouses
  • New AI methods expected to accelerate progress quickly

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