History I: Program it → Plan it → React → Layer it
1961–1999. Four eras, each invented to fix the previous one’s biggest failure, and each leaving a new failure behind. Watch for the pattern: all four ideas are alive inside 2025’s robot foundation models.
4 simulationsReal archival footage14 review cards
Module 1 · 1961≈ 15 min
Program it: Unimate, the first industrial robot
1961 · General Motors, Ewing Township, New Jersey
Unimate
Who
Invented by George Devol (patent filed 1954, “Programmed Article Transfer”); commercialized by Joseph Engelberger’s company, Unimation.
The problem
Die casting: pulling hot metal car parts (door handles and the like) out of a casting machine and dropping them into cooling liquid. Hot, toxic, dangerous work, repeated all day.
The idea
Teach once, replay forever. A worker guided the gripper through the motion; Unimate recorded the positions on a rotating drum memory and replayed them indefinitely. It was a 4,000-pound hydraulic arm with 5 degrees of freedom.
Real world
Unimate even appeared on The Tonight Show with Johnny Carson, where it knocked a golf ball into a cup, poured a beer and waved the bandleader’s baton. Today there are millions of industrial robots, and most still work this way: programmed motions inside a cage.
Solved (vs. a human doing it)
Tireless, precise, repeatable motion in jobs that injured people.
Still couldn’t
Perceive anything. Shift the part 2 cm and Unimate grabs air.
Echo in 2026
Teach-and-replay is the ancestor of imitation learning. The difference is that modern policies learn from many demonstrations and generalize instead of replaying one.
Predict first
The taught spot is fixed. If each part lands within ±4 cm of it, and the gripper only works within ±2.5 cm, what fraction will blind replay pick?
About 2.5 / 4 ≈ 60%. Now run it and see. Then switch on the camera: sensing turns a 60% robot into a ~100% robot, and that’s the whole argument for the next era.
The two answers to “the robot can’t see”Industry: change the world to fit the robot. Fixtures, jigs, conveyors and cages put every part in exactly the same place. This “structured environment” trick still runs most factories, and Day 3 shows Amazon pushing it further. AI researchers: give the robot senses and a brain. That leads to Shakey (Module 3).
Why do most factory robots still not need cameras?
“Structure the world” beats “understand the world” whenever you control the world. Robots struggle exactly where you can’t structure it: homes, warehouses full of varied items, outdoors.
Module 2 · fundamentals≈ 12 min
The servo loop: a 1960s idea inside every 2026 robot
Why this mattersEven “blind” Unimate had feedback inside each joint. Every VLA, from RT-2 to π0.7, outputs targets (“joint 3 to 41°”), and a classic feedback controller running hundreds or thousands of times per second does the actual pushing. Understand this loop and you understand the bottom layer of every robot stack.
A servo measures the error (target − actual) and pushes to shrink it. Three knobs shape how it pushes:
P, the spring: push harder the farther you are from the target. Too strong and it overshoots and wobbles.
D, the damper: resist fast motion, like shock absorbers. It kills the wobble.
I, the grudge: if a small error persists (gravity sagging the arm), slowly build up extra push until it’s gone.
Try:
Press New target a few times with the default knobs. It wobbles: P with almost no D.
Raise D to about 6. The wobble disappears.
Switch on Heavy payload. The arm sags below the target (steady-state error). Gravity is winning a tug-of-war with the spring.
Raise I to about 10. The sag slowly disappears. Push I too high and it overshoots again.
Hit Kick it to see how fast each setting recovers from a shove.
Real worldThe same loop runs cruise control, thermostats, drone stabilizers, hard-disk heads, and the joint controllers underneath a 2025 humanoid’s neural network.
Show me the math (optional)
push = Kp·error + Ki·∫error dt − Kd·velocity
That’s all a PID controller is. The art is in the gains and in what happens at the limits: motors saturate, sensors are noisy, and the loop runs on a clock.
A VLA sends a new joint target every 20 ms, but the joint wobbles around each target. Which knob?
Wobble around a target is under-damping, so add D. (Retraining the VLA wouldn’t help. This is why good low-level control matters even with big models on top.)
Module 3 · 1966–1972≈ 20 min
Plan it: Shakey, the first robot to reason about its actions
1966–1972 · SRI (Stanford Research Institute), Menlo Park
Shakey
Who
Charles Rosen, Nils Nilsson, Peter Hart and colleagues at SRI’s AI Center, funded by DARPA.
The problem
Unimate could only repeat. Could a robot be given a goal (“push the box off the platform”) and work out how by itself?
The idea
Sense → model → plan → act. Shakey turned camera images into a symbolic model of its world (“BOX1 is in ROOM2”), searched for a sequence of actions that reached the goal, then executed it.
Inventions that outlived it
A* search (1968), still how game characters and many route planners find paths. STRIPS (1971), the ancestor of today’s AI planning languages. And the modern form of the Hough transform (1972) for finding lines in images.
Real world
Shakey worked in specially prepared rooms with plain walls and large, simple blocks, and it spent a long time thinking between moves. Life magazine still called it “the first electronic person” (1970).
Solved (vs. Unimate)
Goals and reasoning: the same robot could do many tasks it was never explicitly programmed for.
Still couldn’t
Keep up. Its world model went stale the moment anything moved, its perception was fragile outside tidy rooms, and planning was slow.
Echo in 2026
Google’s SayCan (2022) used an LLM as Shakey’s planner, choosing from a library of robot skills. Modern systems re-plan continuously precisely because of Shakey’s staleness problem.
Play: how A* finds a path
Blind search explores in every direction equally. A* adds a hunch: an optimistic guess of the distance left. Draw walls by tapping, drag S or G to move them, then race the two.
Show me the math (optional)
A* always expands the cell with the smallest
f(n) = g(n) + h(n) cost so far + guess of cost to go
If the guess never overestimates (straight-line or grid distance), A* is guaranteed to find the shortest path. With h = 0 it becomes blind (Dijkstra / breadth-first) search.
1979 · Stanford AI Lab
The Stanford Cart: how slow “think, then act” was
What happened
Hans Moravec’s Cart crossed a chair-filled room on its own in about five hours. It moved in one-metre spurts, then stopped for about ten minutes each time to take pictures, update its map and re-plan.
What Moravec concluded
It is comparatively easy to make computers perform like adults on intelligence tests or checkers, and very hard to give them the perception and mobility of a one-year-old. That is Moravec’s paradox (1988), hard truth #1 from Day 1.
Shakey plans a path, then a person walks into it. What goes wrong, in sense-plan-act terms?
Planning is only as good as the snapshot it planned on. The fix, re-sensing and re-planning constantly, needs fast perception and fast planning. Shakey had neither.
Module 4 · 1986≈ 15 min
React: Brooks says “the world is its own best model”
1986 · MIT AI Lab
Subsumption architecture
Who
Rodney Brooks, “A Robust Layered Control System for a Mobile Robot” (1986).
The problem
Sense-model-plan-act robots were too slow and too brittle for the real, messy world.
The idea
Skip the world model. Stack simple sensor→motor reflexes in layers (“avoid obstacles”, “wander”, “explore”), each a complete behavior on its own. Higher layers can subsume (override) lower ones. Brooks: the world is its own best model, so sense it constantly instead of copying it into a symbolic map.
Real world
Genghis (1989), a six-legged robot that walked over rough ground using simple per-leg reflexes. Brooks co-founded iRobot (1990): its Roomba (2002) brought behavior-based robots into millions of homes, and its PackBots went into the damaged Fukushima reactor buildings in 2011.
Solved (vs. Shakey)
Speed and robustness with cheap sensors and almost no computation.
Still couldn’t
Pursue long-term goals, remember, or plan. Reactive robots get trapped in dead ends (like a U-shaped wall), and every behavior is hand-designed.
Echo in 2026
The fast, learned “System 1” policies inside 2025 humanoids play the role of Brooks’ reflex layers.
Race: Shakey vs. Brooks vs. Hybrid
Three robots, one goal (G), one wandering cat, and a U-shaped trap. Shakey senses, thinks, executes a few metres of its plan blind, then stops to look again. Brooks never plans: it’s pulled toward the goal and pushed away from anything nearby. Hybrid plans a route around the walls, then reacts to the cat on the way.
Predict first
Which robot arrives first? Which one never arrives?
Usually Hybrid wins. Shakey gets there, but slowly: like the real Shakey and the Stanford Cart, it stops every few metres to look again and re-plan, and if the cat wanders into its plan between looks, it bumps. Brooks dodges the cat beautifully, then drives straight into the U and gets stuck: a local minimum, where every nearby move looks worse. Run it a few times; the cat is random (and curious about moving robots).
Why did Brooks get stuck in the U?
Pure reaction is greedy. Escaping a dead end means temporarily moving away from the goal, which needs a plan (or memory). That is exactly the gap the next era filled.
Module 5 · 1990s → 2025≈ 15 min
Layer it, and the 2025 architecture rhyme
1990s · NASA JPL, NASA Ames, and others
Three-layer architectures
The problem
Planners were smart but slow; reflexes were fast but dumb.
The idea
Stack both. A slow deliberative layer plans (minutes), a sequencer or “executive” in the middle picks which skill to run (seconds), and a reactive skill layer runs tight sensor-motor loops (milliseconds). One well-known version is 3T (Bonasso, Firby, Gat and colleagues, 1997).
Real world
In May 1999, NASA’s Remote Agent took control of the Deep Space 1 spacecraft, the first time an AI system controlled a spacecraft without human supervision. It had a planner/scheduler, a “smart executive” that carried out plans, and a diagnosis system that detected faults and recovered, all above the spacecraft’s real-time control software.
Solved (vs. pure planning or pure reacting)
Goals and reflexes in one system.
Still couldn’t
Escape hand-engineering. Every layer, every skill and every interface between layers was written by experts, and the stack still depended on accurate perception and models of the world (Days 3 and 4).
The rhyme: 1997 vs. 2025
Same shape, three decades apart. Colors mark the same role in each stack; labels carry the meaning.
3T (1997)
hand-engineered layers
Planner: decides what to dominutes
Sequencer: picks the next skillseconds
Reactive skills: sensor-motor loopsmilliseconds
Figure Helix (2025)
learned layers, trained end to end, running onboard
S2: 7B vision-language model; understands the scene and the instruction7–9 Hz
S1: 80M visuomotor policy; full upper-body control incl. fingers200 Hz
Motor controllers: servo loops (Module 2)faster
NVIDIA GR00T N1 (2025)
open humanoid foundation model, 2.2B parameters
System 2: Eagle-2 vision-language model10 Hz
System 1: diffusion-transformer action module (flow matching)120 Hz
Motor controllersfaster
Gemini Robotics 1.5 (2025)
an agentic split: orchestrator + action model
Gemini Robotics-ER 1.5: plans multi-step tasks; can call tools like Google Searchslow
Gemini Robotics 1.5 (VLA): turns each step into motions, called by the orchestrator as a toolfaster
Robot controllersfastest
Physical Intelligence’s π0.5 (2025) folds the layers into one network: it first predicts a high-level subtask in words (“pick up the pillow”), then its action expert produces the motor commands for that step.
What changedIn 1997 every layer was hand-written. In 2025 every layer is learned, often from the same data, and sometimes trained jointly (Helix trains S1 and S2 end to end). The shape survived because the physics didn’t change: the world doesn’t wait (hard truth #7).
The systems questions this raises (your research territory)
Rates: how fast must each layer run, and what happens when the slow layer lags behind reality?
Interfaces: what should the layers say to each other: words, latent vectors, waypoints, or skill calls?
Placement: which layer runs on the robot’s own GPU, and which can live in the cloud?
Training: train the layers jointly or separately? Who supplies the data for each?
These return on Days 5–6. Several are open problems in 2026.
Why do 2025 humanoid models split into a slow “thinker” and a fast “doer” instead of running one big model for everything?
It’s the 1990s trade-off reborn: understanding is expensive and slow; reflexes must be cheap and fast. Getting the two to cooperate (rates, interfaces, training) is an active systems problem.
Unimate (GM Ewing Township 1961, drum memory, 5-DoF hydraulic arm, Tonight Show): Wikipedia: Unimate and its cited sources
Stanford Cart timing (five-hour crossing; about ten minutes per one-metre step): Moravec’s own account; his 1980 thesis
Shakey, its inventions, and Life’s 1970 “first electronic person”: Wikipedia: Shakey the robot; Hart, Nilsson & Raphael (1968) on A*; Duda & Hart (1972) on the Hough transform