The Paradigm Shift Toward Embodied AI

Seminar for Comprehensive Competency Cultivation
HSS118 · W02 · DGIST

Yunsung Lee

2026-09-04

About me

Yunsung Lee
Yunsung Lee

WoRV: World Models for Robotics and Vehicle Control
Robotics research lab at Maum.AI

Co-instructor: Chris Choi, CEO of Maum.AI
closing keynote on Dec 4

Maum AI

Picking up from last week

  • Last Friday (Aug 28): Special lecture on AI
  • This course: one semester on AI that gets a body

The semester in one picture

Semester timeline, week 02 highlighted AUG SEP OCT NOV DEC today Sep 4 Sep 11 holiday Oct 2 holiday report Oct 23 Nov 6 Nov 20 Nov 27 Dec 4 essay Lecture Guest lecture Keynote Leadership talk holiday / report / essay week

Attendance: two numbers

  • In person: sign the sheet; the TA spot-checks at random
  • Attendance and participation: 40% of the grade
  • Three absences: F
  • Remote Sep 11 / Nov 6: DGIST login records attendance

Official notices, assignments, deadlines and grades live on the DGIST LMS.

Essays: submitting is what counts

  • Eight talks: five lectures, two guests, one closing keynote
  • Reflection after each talk: 50%; submission counts; three missing = F
  • Final synthesis essay: 10%; prompts and deadlines on the LMS

Official notices, assignments, deadlines and grades live on the DGIST LMS.

Ask anytime

QR code for the question wall
Wooclap code ZNQKOGFapp.wooclap.com/ZNQKOGF

Questions go on the wall any time; we answer them in the last 10 minutes.

Autonomous or teleoperated?

Hands up: did Atlas move on its own?

Boston Dynamics says Trained in simulation.
Runs on its own. Atlas · May 2026 · autonomy claimed

Never said “No teleoperation.”

For comparison This is teleoperation →

Embodied Avatar: full-body teleoperationUnitree · Nov 2025 · teleoperated · 1x speed

Moravec’s paradox (1988)

The hard problems are easy.
The easy problems are hard.

Chess fell in 1997. An AI can write your essay.
Fetching a drink held out until now.

Hans Moravec, Mind Children (1988) - paraphrased.

AI is getting a body: from AI that talks to AI that acts.

Today’s route

  1. The problem - why easy is hard (done)
  2. The history - how AI learned to see and talk, 2012-2022
  3. The turn - why action is next, and why now
  4. The map - where this course goes

1

How AI Learned to See and Talk

2012 - 2022

2012 · AlexNet: Machines Learn to See

ILSVRC top-5 error by year Bar chart showing top-5 error falling from 28.2 percent in 2010 to 3.57 percent in 2015. Before 2012 the yearly gain was 2.4 points; AlexNet in 2012 cut 9.4 points in one year, and the deep-learning era after it passed the human reference of 5.1 percent within three years. top-5 error (%) - lower is better 28.2% 25.8% 16.4% 11.7% 6.7% 3.57% −2.4 pt −9.4 pt 2010 2011 2012 2013 2014 2015 deep learning: below human in three years human ~5.1% AlexNet - deep learning arrives
  • ImageNet: 1.2 million labeled photos, plus two gaming GPUs
  • No hand-written rules - the network learned from examples
  • Within three years it beat a human at this benchmark

Metric: ILSVRC top-5 error · Data: Russakovsky et al. (2015); 2015 point: He et al. (2016), ensemble · Try a CNN yourself

2017 · One architecture for everything

You don’t need to read this diagram. Just remember the name: Transformer.

Transformer architecture from Vaswani et al. (2017)

Figure: Vaswani et al. (2017), “Attention Is All You Need” · Interactive: poloclub.github.io/transformer-explainer

2018-2020 · GPT: One Simple Game

Guess the next word. That was the whole training.

“Scientists discovered unicorns in the Andes
that spoke perfect ____” English.

GPT-2 continues (2019):
“The scientist named the population, after their distinctive horn, Ovid’s Unicorn.”

Fluent is not the same as true.

Sample: OpenAI (2019), “Better language models and their implications”

2020 · Bigger kept getting better

Three concept panels for AI scaling laws Three simple panels show error descending in straight lines as compute, data, and model size increase. error more compute more data bigger model
  • Bigger model + more data + more compute = predictably better
  • A straight line, again and again - no ceiling in sight
  • That one chart became the industry’s playbook: scale up

Concept chart after Kaplan et al. (2020), Fig. 1

2022 · Language Meets Images

One sentence in, a painting out: “a fox in a field, in the style of Monet”

2021 · DALL-E 1

2022 · DALL-E 2

2025 · the tool you use today

Source: OpenAI (2022, 2025)

2022 · Talking AI becomes a habit

100 million users in two months.

Talking AI stopped being a demo and became a habit.

Estimate: UBS / Similarweb, reported by Reuters (Feb 2023)

Foundation Models: Train Once, Do Many Things

  • One model from varied data
  • Adapt it to many tasks
  • That is a foundation model

Foundation model pipeline from Bommasani et al. (2021)

Figure: Bommasani et al. (2021), Fig. 2 - CC BY 4.0

Meanwhile, Robots…

Atlas | Partners in ParkourBoston Dynamics · Aug 2021 · autonomous · 1x speed
  • Model-based control made dynamic whole-body motion possible.
  • Learning systems now build on that engineering foundation.

Two Waves, One Wall

Wave 1 · Trial and error
Wave 2 · Imitation
QT-Opt robot farm
Robot farm - Google, 2018
RT-1 task montage
Thousands of demonstrations - Google, 2022
Robot hand holding a Rubik's Cube
Robot-hand dexterity - OpenAI, 2019
ALOHA dual-arm manipulation
Two arms copy human hands - Stanford, 2023

Both waves hit the same wall: generalization.

Sources: Kalashnikov et al. (2018) · OpenAI (2019) · Brohan et al. (2022) · Zhao et al. (2023)

What Was Still Missing

The robots of that decade were missing three things.

  • Common sense - everyday knowledge of how the world works
  • Language - you could not just tell a robot what you wanted
  • The open world - anything beyond the scripted lab task

By 2022, talking AI suddenly had all three.
What happens when you give it a body?

2

The Turn

Why action is next - and why now

Same recipe, new output

Language models progress from words to actions A three-stage rail compares the inputs and outputs of a Language Model, a Vision-Language Model, and a Vision-Language-Action model. Language Model (LM) 1 INPUT text in OUTPUT next word out Vision-Language Model (VLM) 2 INPUT image + text in OUTPUT words about the world out Vision-Language-Action (VLA) 3 INPUT observation + instruction in OUTPUT next action out
  • Language models learn one trick: predict the next word.
  • Give the model eyes, and the same trick describes images (VLM).
  • Give it a body, and the same trick predicts the next action (VLA).

The only formula this course needs

\pi(a \mid o, l)

  • a - the action the robot takes
  • o - what the robot observes
  • l - the instruction, in language
  • \pi - the rule that picks the action (researchers call it a policy)

It comes back every week.

RT-2: web knowledge moves a robot arm

RT-2: pick up the extinct animalGoogle DeepMind · Jul 2023 · autonomous · 1x speed
  • “Pick up the extinct animal” - it picks the dinosaur.
  • The command never appears in its robot training data.
  • Web knowledge starts moving a robot arm.

What does Embodied AI mean?

Embodied AI closes the loop in one body: seeing, understanding, deciding, and acting.

Embodied AI ≈ Physical AI industry term Robot Learning the academic field RFM · Robot Foundation Models

From here on, we use the term Embodied AI.

Every new model bets on one of three axes

Three axes of generalization Three panels show task, environment, and embodiment generalization with one example model for each axis. TASK new verb, same room fold → roll EXAMPLE RT-2 ENVIRONMENT same verb, new room kitchen/day → living room/night EXAMPLE π0.5 EMBODIMENT same brain, new body parallel-jaw → dexterous hand → humanoid EXAMPLE GR00T N1

The data barrier

The data barrier on a log scale Horizontal log-scale bars compare web text, web images, and robot teleoperation. The difference spans many orders of magnitude. many orders of magnitude Web text ~10^13 words Web images ~10^7 to 10^9 Robot teleoperation ~10^4 hours 10^3 10^4 10^5 10^6 10^7 10^8 10^9 10^10 10^11 10^12 10^13

Action is not on the internet.

Log scale; units differ by type (words / images / hours) - read orders of magnitude, not counts.

π0 folds laundry, uncutPhysical Intelligence · Oct 2024 · autonomous · 1x speed

π0.5 works across unseen homes (10x)Physical Intelligence · Apr 2025 · autonomous · 1x speed

Helix: two robots, one brain

Helix: two robots put away groceriesFigure · Feb 2025 · autonomy not stated · 1x speed
  • Two Figure robots share one learned policy (Feb 2025).
  • By 2026, Figure describes Helix 02 as “fully autonomous, not teleoperated.”

Sources: Figure (2025, 2026). The autonomy wording belongs to the 2026 Helix 02 announcement.

CMG 2026 Spring Festival Gala (Unitree)Unitree · Feb 2026 · autonomy claimed · 1x speed

Atlas: The Beginning of Your Tomorrow (CES 2026)Boston Dynamics · Jan 2026 · autonomy not stated · 1x speed

NEO, the home robot - launch demos were teleoperated1X · Oct 2025 · autonomy not stated · 1x speed

The 2023-2026 unlock

RT-2 robot
Jul 2023
RT-2
OpenVLA robot
Jun 2024
OpenVLA
Physical Intelligence pi zero
Oct 2024
π0
Figure Helix robots
Feb 2025
Helix
Gemini Robotics
Mar 2025
Gemini Robotics
NVIDIA GR00T N1
Mar 2025
GR00T N1
Physical Intelligence pi zero point five
Apr 2025
π0.5
Gemini Robotics 2
Jul 2026
Gemini Robotics 2

Everything you just watched happened in 36 months.

Sources: Google DeepMind · OpenVLA, Kim et al. (2024) · Physical Intelligence · Figure · NVIDIA

The pull now comes from industry, not academia

Capital
Figure $1B+ at $39B; Skild $1.4B at $14B+; NEURA up to $1.4B.
Production
Unitree reports 5,500+ humanoids shipped in 2025; its $904M IPO rose 460% on debut. It was mainland China’s first listed humanoid maker.
Platforms
NVIDIA’s open GR00T reference humanoid and Google DeepMind’s whole-body Gemini Robotics 2 arrived in 2026.

Capital: Figure, Sep 16 2025 · Skild, Jan 14 2026 · NEURA/CNBC, Jun 10 2026
Production: Bloomberg/Unitree, Aug 19 2026 (shipments company-reported)
Platforms: NVIDIA, Jun 1 2026 · Google DeepMind, Jul 30 2026

One case from where I work

CostNav: navigation judged by cost, in simulationWoRV · Nov 2025 · autonomy not stated · 1x speed
In-house dual-arm manipulationWoRV · Mar 2026 · autonomous · 1x speed

Left: can a robot’s route earn its keep? Right: can two arms do real work?

From demo to deployment

  • Precise contact - buttons, cables, glassware.
  • Long horizons - an hour of chores without a reset.
  • Unfamiliar bodies - policies still cling to their own hardware.
  • Speed and safety - human pace, around actual humans.

The next question: what capability can scale?

3

The map

Where this course goes

The semester, again

Semester timeline, week 02 highlighted AUG SEP OCT NOV DEC today Sep 4 Sep 11 holiday Oct 2 holiday report Oct 23 Nov 6 Nov 20 Nov 27 Dec 4 essay Lecture Guest lecture Keynote Leadership talk holiday / report / essay week

Today set the question. The rest of the semester answers it.

First: how robots learn

DreamZero World Action Model
Sep 11 · Video Foundation Models and World Models
Guest: Hyeongmin Lee · Robots that learn by watching video.
Open X-Embodiment data overview
Oct 2 · Action Data
Where robot experience comes from.
Diffusion Policy gradient field
Oct 23 · Learning Methods
How demonstrations become skills.

Sources: DreamZero (NVIDIA, 2026) · Open X-Embodiment Collaboration (2023) · Chi et al. (2023)

Then: from policy to physical motion

VGGT photos to 3D reconstruction
Nov 6 · Spatial AI to Physical AI
Guest: Sunghwan Hong · Machines that know where things are.
ALOHA teleoperation hardware
Nov 20 · Robot Hardware Special
Control and mechanics beneath learned policies.

Sources: Wang et al. (2025) · Zhao et al. (2023)

Finally: safety and industry

Gemini Robotics 2 Safety Evaluations cover
Nov 27 · Ethics and Safety
What robots must never do, and who checks.
Figure 03 humanoid
Dec 4 · Closing Keynote
Chris Choi · Physical AI as a megatrend: the view from industry.

Sources: Google DeepMind (2026) · Figure (2025)

Three takeaways

  1. Same recipe, new output: action.
  2. Data is the bottleneck: action is not on the internet.
  3. The next frontier: generalize and scale.

Your questions

QR code for the question wall
Wooclap code ZNQKOGFapp.wooclap.com/ZNQKOGF