Three AIs
walk into
a bar.

Marcin Wysocki
machinekind.ai
whoami

Marcin Wysocki.

Marcin at a table with a laptop, a small robot arm and robot parts with the Machinekind logophoto
assets/me.jpg
machinekind.ai
assets/cool/droids.jpg
assets/cool/k2so.jpg
assets/cool/ironman.jpg
assets/cool/bb8.jpg
Zytonits (CC BY-SA 2.0) · Fred Cherrygarden (CC BY-SA 4.0) · dambranslv (CC BY 2.0)

Why now?

Humanoids + robot foundation models · $B a year
All robotics · $B a year · their share in red
Sources: F-Prime (US, Europe, Israel) · Dealroom, Crunchbase (global)
machinekind.ai

Seen this curve before?

Robots · humanoids + robot foundation models · $B a year
LLMs · generative AI · $B a year · 24× the scale
Sources: F-Prime (US, Europe, Israel) · Stanford AI Index 2026 · Dealroom: humanoids, AI rounds
machinekind.ai
Claude Opus 5.5 paints in watercolor · @dhvanil
machinekind.ai
RL SLAM VLM IK PPO JEPA GAE VLA DOF WAM MPC ACT Sim2Real BC LLM π0 RT-2 GR00T IL MJX PD ViT TD-MPC DR IMU LBM SAC CNN VQA MLP

What is
a robot?

machinekind.ai
Definition

A machine controlled by a computer that is used to perform jobs automatically.

machinekind.ai
assets/real/industrial.jpg
assets/real/cnc.jpg
assets/real/roomba.jpg
assets/real/thermomix.jpg
Kārlis Dambrāns from Latvia (CC BY 2.0) · Jan Hagelskamp1 (CC BY 4.0)
assets/now/atlas.jpg
assets/now/optimus.jpg
assets/now/g1.jpg
assets/now/spot.jpg
assets/now/anymal.jpg
Damian B Oh (CC BY-SA 4.0) · Steve Jurvetson (CC BY 2.0) · Sayanesy (CC0) · Jonte (CC BY-SA 4.0) · ANYbotics (CC BY-SA 4.0)
Zoom in

Robot is a system.

1 2 2 3 4
Hardware
motors1
sensors2
electronics3
battery4
Software
motor controllers
control · telemetry
AI · reasoning
machinekind.ai

What does this system need to fetch us a beer?

machinekind.ai
Three tasks

Locomotion
Navigation
Manipulation.

machinekind.ai
Three tasks

Locomotion.
Navigation
Manipulation

machinekind.ai
How do you make it walk

Three ways.

Write the physics
IK + MPC
Let it try
RL + PPO
Show it how
Imitation
The first needs no AI · the other two need a simulator
machinekind.ai
The example

Wojtek.

RL + PPO
50 Hz
runs on a Raspberry Pi
video
assets/wojtek.mp4
machinekind.ai
What a policy is

A small net, fifty times a second.

JOINT ANGLES×12JOINT SPEEDS×12LAST ACTION×12COMMAND×4 Small MLP 12 target angles 50 Hz Motor drives 512 · 256 · 128~190K WEIGHTS ONE PER MOTOR 40 NUMBERS INCOMMAND = FORWARD · SIDEWAYS · TURN · HEIGHT 12 NUMBERS OUTPD → TORQUE
No camera · no IMU · no map · no idea where it is
machinekind.ai
Three tasks

Locomotion
Navigation.
Manipulation

machinekind.ai
Getting to the bar

Four questions, one needs AI.

code
Where am I?
code
What’s in the way?
code
How do I get there?
AI
Which thing is the bar?
machinekind.ai
Four tasks

Reasoning.
Locomotion
Navigation
Manipulation

machinekind.ai
Reasoning · the model plans the work · video by @dhvanil
machinekind.ai
VLM · vision-language model

Image and words in, text out.

“walk to the purple pillar” Vision encoder Language model {"type": "goal", "point_2d": [670, 490], "label": "purple pillar"} Depth + controller CAMERA FRAME · PATCHESINSTRUCTION VIT + PROJECTORIMAGE TOKENS TOKENIZERTEXT TOKENS TEXT · ONE JSON OBJECTMOTION
A VLM only ever writes text · other code turns it into motion
machinekind.ai
Problem?

Slow AF.

modelfastest settinganswers per secondthinking hard
Gemini 4 Argon––41 s
Claude Fable 5.28.5 s0.12 Hz267 s
Claude Opus 5.54.8 s0.21 Hz343 s
GPT-6 Astra3.6 s0.27 Hz229 s
Qwen3-VL 30B the one on Wojtek2.6 s0.38 Hz–
Gemma 4 26B1.7 s0.60 Hz27 s
Nemotron Nano 12B VL1.5 s0.66 Hz–
Qwen3 Omni 30B1.2 s0.86 Hz26 s
Llama 4 Scout1.1 s0.88 Hz–
Phi-4 Multimodal0.42 s2.4 Hz–
Wojtek’s walking policy0.02 s50 Hz–
Time to the first word of the answer · cloud APIs · Artificial Analysis, Nov 2026
machinekind.ai
VLA · vision-language-action model

One step further.

“put strawberry into the correct bowl” Vision encoder Language model 1 128 91 241 5 101 127 217 1128 91 2415 101 127217 STOP FLAGΔ POSITIONΔ ROTATIONGRIPPER Arm controller CAMERA FRAME · PATCHESINSTRUCTION VIT + PROJECTORIMAGE TOKENS TOKENIZERTEXT TOKENS 8 ACTION TOKENS · RT-2’S OWN EXAMPLEMOTION
RT-2 · 2023 · 55B · 1–3 HzOpenVLA · 2024 · 7B · 6 Hz
The tokens it writes are the action · the first ones were still slow
machinekind.ai
Fast and slow

Slow brain, fast hands.

modelthinksone thought givesmovesthe fast part is
OpenVLA 20246 Hz1 move6 Hznone
GR00T N1 202510 Hz16 moves120 Hzreplay of the batch
Gemini Robotics 2025~4 Hza batch of moves50 Hzreplay of the batch
π0.5 2025–50 moves50 Hzreplay of the batch
Helix 20257–9 Hza goal, not moves200 Hza 2nd network that keeps looking
Either way every move comes out of the VLA · ~ = worked out · – = not published
machinekind.ai
Four tasks

Reasoning
Locomotion
Navigation

Manipulation.

machinekind.ai
Whole-body VLAs · November 2026

Feet to fingertips.

modelrobothow the body is driventheir headline
Gemini Robotics 2.5 Google · Nov 2026Apollo 2a VLA for the whole body, 22-DoF hands96% · factory kitting pilot, arms + hands
Helix 02 Figure · Jan 2026Figure 03three networks: planner · joints at 200 Hz · balance at 1 kHz4 min, 61 steps, no resets
GR00T N1.7 NVIDIA · 2026Unitree G1the VLA writes tokens, a learned controller moves the jointsopen weights · 28 Hz on a Jetson
Each number is the maker’s own, on its own task
machinekind.ai
Where the data comes from

Show it how.

video · assets/data/teleop.mp4
Teleoperation35 demos an hour
video · assets/data/hand-held-gripper.mp4
Hand-held gripper111 demos an hour
umi-gripper.github.io
machinekind.ai
World model · Tesla · Oct 2025

Same scene, rewritten.

video · assets/tesla/world-sim-cut-in.mp4
FSDleft: original · right: the model makes a car cut in
video · assets/tesla/world-sim-optimus.mp4
Optimusgenerated, same simulator
Tesla · FSD clip · Optimus clip
machinekind.ai
WAM · world action model

See, imagine, act.

“put the banana on the plate” Video encoder Video model Robot CAMERA FRAMES · UP TO 6.6 SINSTRUCTION VAEVIDEO LATENTSJOINT STATE TEXT ENCODERTEXT EMBEDDINGS VIDEO DIFFUSION · 14B NEXT FRAMES · IMAGINEDTHEN REPLACED BY THE REAL ONES 48 MOVES · 1.6 S AT 30 HZMOTION
DreamZero · NVIDIA · 2026 · 14B · 7 HzGR00T N2 · built on DreamZero · due by end of 2026
One model predicts the video and the moves together · DreamZero paper
machinekind.ai
Scorecard

November 2026.

taskstatusin one line
reasoningslowgood plans, a few seconds each
navigationgluemaps and planners drive; AI only names the goal
locomotionrealdeployed, cheap, boring in the best way
manipulationpilotfirst factory pilots; narrow, hungry for data
world models, WAMslabon real robots in the lab; first product promised for 2026
machinekind.ai
01 / 29