// DECK 13 / TECH-ROBOTICS / 2026

ROBOTICS /
Machines
that move

FROM INDUSTRIAL ARMS  →  HUMANOIDS

A short tour of the field: from R.U.R. on a 1920s stage to bipeds walking onto factory floors a century later.

01 / ETYMOLOGY

The word.

The term robot enters language in 1920, in Karel Čapek's play R.U.R. — Rossum's Universal Robots. Coined by his brother Josef, it derives from Czech robota: forced labor, drudgery.

The play's robots are not metal. They are mass-produced synthetic workers who eventually rise against their makers — the template for a century of anxiety about machine labor.

> PLAY : R.U.R.
> AUTHOR : KAREL ČAPEK
> PREMIERE : PRAGUE, 25 JAN 1921
> ETYMOLOGY : robota (Cz.) — forced labor
02 / FICTION SHAPES POLICY

Three Laws.

Isaac Asimov, Runaround (1942). A fictional safety framework embedded in every positronic brain — and a permanent reference point for real-world AI ethics debates.

Law 01

A robot may not injure a human being or, through inaction, allow a human being to come to harm.

Law 02

A robot must obey orders given it by human beings, except where such orders conflict with the First Law.

Law 03

A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.

Asimov spent decades writing stories about how cleverly these laws fail. Real robots run on ROS, not ethics — but the framing stuck.

03 / 1961

Unimate at GM.

The first industrial robot enters service at General Motors' Inland Fisher Guide plant in Ewing Township, NJ. It was a 4,000-lb hydraulic arm built by George Devol and Joseph Engelberger, lifting hot die-cast parts that would have maimed humans.

> UNIT : UNIMATE 1900
> YEAR : 1961
> SITE : GM TRENTON / DIE CASTING
> PAYLOAD : ~225 kg
> CONTROL : MAGNETIC DRUM, HYDRAULIC
FIG.01 — UNIMATE / 1961
04 / SCALING THE ARM

The industrial era.

By the 1980s, Japan, Germany, and Switzerland turned the industrial arm into a global commodity. Six revolute joints, repeatable to fractions of a millimeter, bolted to factory floors worldwide.

FANUC

Yamanashi, Japan. Yellow arms. ~750k installed.

ABB

Zürich. White arms. IRB lineage since 1974.

KUKA

Augsburg, Germany. Orange arms. KR series.

> AXES : 6 (BASE/SHOULDER/ELBOW/WRIST×3)
> REPEAT : ±0.02 mm
> FLEET : ~3.5 M operational worldwide (2024)
J1 J2 J3 FIG.02 — 6-AXIS ARTICULATED ARM
05 / CANONICAL LOOP

Sense. Plan. Act.

Every robot, from a Roomba to Atlas, runs some version of this loop. The 1980s "subsumption architecture" debate was about whether you could skip the plan stage; modern systems blend reactive and deliberative layers.

  • SENSE — cameras, LIDAR, IMU, force/torque, joint encoders.
  • PLAN — state estimation, motion planning, task graphs.
  • ACT — joint torques, end-effector commands, locomotion gaits.

Loop frequency matters: 1 kHz at the joint, 30–200 Hz at perception, ~10 Hz at task. Latency is the hidden enemy.

SENSE PLAN ACT FIG.03 — CONTROL LOOP
06 / NAVIGATION

SLAM: build a map of a place you've never been.

Simultaneous Localization and Mapping. The chicken-and-egg problem at the heart of mobile robotics: to know where you are, you need a map; to make a map, you need to know where you are. SLAM solves both at once, probabilistically.

  • EKF-SLAM (1990s) — extended Kalman filter, gaussian beliefs, doesn't scale.
  • FastSLAM / GraphSLAM (2000s) — particle filters, pose graphs, loop closure.
  • Visual SLAM — ORB-SLAM, DSO; cameras only, runs on a phone.
  • NeRF / 3D Gaussian SLAM (2023+) — dense photoreal reconstruction in the loop.
> KEY OPS : feature extract → data assoc → optimize pose graph → close loop
> FAILURES : kidnapped robot, perceptual aliasing, dynamic scenes
07 / DYNAMIC LOCOMOTION

Boston Dynamics.

Spun out of MIT's Leg Lab in 1992 under Marc Raibert. They proved a robot could fall and not fall — that controlled instability beat statically stable plodding. The viral videos pulled the field forward by years.

BigDog · 2005

DARPA-funded gas-powered quadruped. Shoved on ice, recovers. The "do not anger it" video.

Atlas · 2013→

Hydraulic, then electric (2024). Backflips, parkour, picking parts on a mock factory floor.

Spot · 2019

First commercial product. Inspections at oil rigs, construction sites, hospitals. ~$75k.

→ search: boston dynamics atlas (YouTube)

08 / THE HARD PART
⚠ Unsolved

Manipulation is still hard.

Walking and driving are largely solved. Picking up a strawberry without crushing it, opening a ziplock bag, threading a cable — these remain open problems. The reasons are physical, not algorithmic.

  • Contact dynamics — friction, deformation, slip. Hybrid systems are hard to plan through.
  • Tactile sensing — GelSight, BioTac give rich signals; few hands have them at scale.
  • Dexterity — the human hand has 27 DoF. Most robot hands have 4–16, with weak fingertips.
  • Dataset gap — internet has billions of images but few demonstrations of contact.
> HUMAN HAND : 27 DoF, ~17,000 mechanoreceptors
> SHADOW HAND: 24 DoF, ~$100k+
> GAP : not joint count — sensing + control bandwidth
09 / 2022 — ?

The humanoid wave.

A surge of well-funded entrants betting that human-shaped robots can drop into existing human environments — warehouses, homes, factories — without retooling.

  • Tesla Optimus — vertically integrated, leveraging FSD stack.
  • Figure — partnered with BMW, OpenAI ties, Bay Area.
  • 1X (Norway) — NEO Beta soft-shell home humanoid.
  • Apptronik — Apollo, Mercedes pilots.
  • Unitree (China) — G1, H1; aggressive pricing under $20k.

As of 2025: pilots in warehouses. Unit economics, MTBF, and safety certification are the gating questions, not motion.

FIG.04 — BIPEDAL HUMANOID
10 / WHAT CHANGED
2022 → 2025

The AI brain.

For 50 years, robotics control was hand-engineered. The recent shift: large pretrained models — vision, language, and action — fine-tuned on robot data, generalize across tasks the way LLMs generalize across text.

  • RT-1 / RT-2 (Google, 2022–23) — transformer policies on real robots.
  • OpenVLA (2024) — open-source vision-language-action, 7B params, fine-tunable.
  • π0 / π0.5 (Physical Intelligence, 2024–25) — flow-matching action heads.
  • Gemini Robotics (2025) — multimodal foundation models targeting embodiment.
> INPUT : RGB frames + language instruction
> OUTPUT : end-effector pose / joint deltas @ 5–50 Hz
> UNLOCK : zero-shot tasks, language-conditioned policies, sim-to-real transfer
11 / DEPLOYMENT

Where the money goes.

WAREHOUSES

Amazon, Symbotic, Locus. Picking, sorting, end-of-line. Highest-volume near-term market.

ELDERCARE

Japan + Europe demographics. Lift assist, fall detection, social companions. Slow regulatory path.

SURGERY

Intuitive's da Vinci has done 14M+ procedures. Tele-operated, not autonomous — yet.

AGRICULTURE

Tractors steer themselves; weeders, pickers, dairy bots fill labor gaps.

LOGISTICS

Autonomous trucks, last-mile delivery rovers, sidewalk bots. Regulation-bound.

DEFENSE

Drones, ground vehicles, perimeter security. The fastest-deploying segment, and the one with the most ethical weight.

12 / CLOSING

References & further viewing.

A starter set. None are exhaustive; all are useful.

  • Books — Siegwart & Nourbakhsh, Intro to Autonomous Mobile Robots; Murray, Li, Sastry, A Mathematical Introduction to Robotic Manipulation.
  • Open source — ROS / ROS2, MoveIt, Drake, Isaac Sim, MuJoCo.
  • Courses — Underactuated Robotics (Russ Tedrake, MIT); Modern Robotics (Kevin Lynch, Northwestern).
  • Reports — IFR World Robotics annual; Robot Report.

// VIDEO SEARCHES

END OF DECK / 13 SLIDES / SAFETY: KEEP CLEAR