From Shakey to Humanoids: The 60-Year Quest for Embodied Machine Intelligence

While computer scientists in the 1960s believed that building a chess-playing program would be the hardest intellectual challenge and that giving a machine arms and legs would be a trivial engineering follow-up, reality delivered the opposite verdict. Known as Moravec's Paradox, the artificial intelligence community discovered that reasoning about abstract logic requires astonishingly little computation, whereas walking down a flight of stairs or grasping an unmodeled coffee cup requires billions of real-time calculations per second.

The journey from the jerky, room-scale computers of the late 1960s to modern dynamic humanoids walking factory floors spans six decades of hard-won breakthroughs across sensors, dynamic equilibrium, and neural policy synthesis.

1. Shakey at SRI (1966–1972): The First Autonomous System

Constructed at the Stanford Research Institute (SRI) under the leadership of Charles Rosen, Nils Nilsson, and Bertram Raphael, Shakey the Robot was the first mobile machine to integrate perception, automated reasoning, and physical actuation into a closed loop.

Shakey's physical hardware consisted of a motorized two-wheel differential drive base, an onboard television camera, an optical rangefinder, and whisker contact sensors. However, because onboard computers of the late 1960s were too massive and hot to ride on a mobile chassis, Shakey transmitted sensory telemetry via analog radio links to an SDS-940 time-sharing mainframe computer.

Shakey earned its nickname from the intense mechanical shaking of its chassis when stepping between movements. A single plan execution often took several hours as the mainframe processed grainy vidicon camera frames.

2. The Stanford Cart (1979): The Compute Barrier of Early Stereo Vision

At Stanford University, Hans Moravec turned attention to outdoor autonomous mobility with the Stanford Cart. Unlike Shakey, which navigated structured laboratory corridors with flat planar lighting, the Cart faced arbitrary shadows, uneven grass, and variable sunlight.

To perceive 3D depth without active laser rangefinders, Moravec mounted a single TV camera on an electric slider rail. The camera moved across nine equidistant horizontal positions, capturing nine overlapping frames to compute depth via parallax feature matching. However, calculating cross-correlations across nine 256x256 pixel images on a DEC KL-10 mainframe took roughly 15 minutes per step:

[Stop Cart] -> [Slide Camera 9 Positions] -> [Transmit Video to KL-10]
           -> [15 Minutes Correlation Compute] -> [Roll 1 Meter Forward] -> [Repeat]

Moravec's Cart conclusively proved that vision-based mobile robotics was fundamentally compute-starved: real-time locomotion would remain impossible until microprocessors achieved orders of magnitude higher arithmetic throughput.

3. The Subsumption Rebellion (1986): Brooks & Behavior-Based Robotics

By the mid-1980s, the classical robotics paradigm—Sense → Model → Plan → Act (SMPA)—had ground to a halt. Robots spent 99% of their operating time constructing fragile internal symbolic world models that became stale the moment an obstacle moved.

In 1986, Rodney Brooks of the MIT AI Lab published his radical manifesto, "A Robust Layered Control System for a Mobile Robot," followed by "Elephants Don't Play Chess" (1990). Brooks argued that the world itself is its own best model. He introduced the Subsumption Architecture:

• Eliminated centralized world models and symbolic planning completely.
• Built robots as asynchronous, layered collections of simple behavioral finite-state machines.
• Lower layers (e.g., Level 0: Avoid Collisions via bumpers) could subsume or suppress higher layers (Level 1: Wander, Level 2: Explore).

Walking hexapods like Genghis demonstrated that lifelike, insectoid navigation across rough terrain could be executed using tiny 8-bit microcontrollers with zero planning overhead. This philosophy directly led to the commercial success of the iRobot Roomba in 2002.

4. Honda ASIMO (2000): Dynamic Bipedal Walking & Zero Moment Point (ZMP)

While MIT focused on reactive insect mobility, Honda Motor Company embarked in 1986 on a top-secret humanoid robotics program. The fundamental physics challenge of two-legged walking is that a biped spends most of its gait in unstable equilibrium.

Honda's engineers (developing the E-series experimental bipeds and P-series prototypes) unlocked dynamic walking by operationalizing Miomir Vukobratović’s Zero Moment Point (ZMP) theory:

Unveiled in 2000, ASIMO (Advanced Step in Innovative MObility) stood 120 cm tall, weighed 43 kg, and utilized custom high-ratio harmonic drive electric actuators. ASIMO stunned the world by walking smoothly at 2.7 km/h, climbing stairs autonomously, and dynamically jogging at 6 km/h with both feet leaving the ground simultaneously.

5. Boston Dynamics Atlas: High-Pressure Hydraulics & Dynamic Parkour

While ASIMO achieved stable quasi-static walking on smooth floors, pushing it or having it step onto slippery ice caused joint torque saturation and catastrophic falls. ZMP was fundamentally conservative: it required keeping feet flat against the ground.

Marc Raibert’s Boston Dynamics took an entirely different mechanical approach with Atlas (unveiled in 2013 for the DARPA Robotics Challenge). Instead of electric gear motors, Atlas employed custom, ultra-high-pressure hydraulic actuators operating at 3,000 PSI (21 MPa):

• Immense Power Density: Compact hydraulic cylinders delivered torque densities unattainable by conventional electric motors, allowing rapid limb accelerations.
• Model Predictive Control (MPC): Instead of enforcing flat soles, Atlas utilized continuous trajectory optimization over full rigid-body dynamics, treating balance as dynamic recovery through stepping and momentum transfer.

Atlas mastered gymnastics, backflips, parkour, and multi-obstacle leaping. However, hydraulic systems suffered from severe commercial liabilities: fluid leaks, high thermal dissipation, intense acoustic noise, and excessive battery drain (limiting autonomy to under 45 minutes).

6. The Modern Electric Humanoid Fleet: 2024–2026

Between 2024 and 2026, the robotics landscape experienced an unprecedented commercial convergence. Improvements in high-torque quasi-direct drive (QDD) BLDC motors, cycloidal/planetary gearboxes, high-density lithium batteries, and vision-language-action foundation models rendered hydraulic systems obsolete.

Boston Dynamics retired its hydraulic Atlas in April 2024, replacing it with a fully electric humanoid featuring 360-degree rotational joint freedom. Concurrently, commercial humanoid ventures entered automotive manufacturing facilities:

7. 60 Years of Engineering Lessons

Reviewing sixty years of embodied intelligence reveals four foundational truths that govern every physical machine:

1. Perception Must Be Bound to Latency: A robot that perceives the world with 500ms latency cannot balance. Cognitive semantic reasoning can run slowly (1-5 Hz), but motor stabilization must execute deterministically at 100 Hz to 1,000 Hz.
2. Hardware Limits Dictate Algorithmic Feasibility: Shakey was limited by vacuum tubes; the Stanford Cart by memory bandwidth; ASIMO by motor power density. Today's neural revolution was unlocked by GPU-parallelized simulation (MuJoCo, Isaac Sim).
3. Sutton's Bitter Lesson Applies to Physical AI: Hand-crafted kinematic heuristics and manual inverse kinematics are consistently superseded by general learned policies trained on massive simulation data.
4. The Reality Gap is Overcome Through Randomization: No simulator is perfect. Humanoids achieve physical robustness not through mathematical perfection, but by training across millions of aggressively randomized simulated physical worlds.