Data operations
The training mix is where advantage accumulates.
Web priors create breadth, simulation creates cheap practice, and real demonstrations create physical truth. Advantage comes from making the three compound.
Read the data thesis
Strategic research database
Berlin · 2026
RI / 001 / 010
Robotics intelligence
for decisions that
have to touch reality.
A working database of market maps, technical mental models, and strategic implications for embodied AI.
Current thesis
The advantage is moving from impressive demos to repeatable
data, deployment, and learning systems.
Open the index
Briefings
10
Platforms tracked
72
Strategic lenses
05
Archive since
06.02.26
01 / Signals in view
What the field is telling us now.
System architecture
Intelligence is splitting by clock speed.
Slow general reasoning is commoditizing. Fast, embodied control and the interface between planning and reflex remain defensible product territory.
Inspect the architectureMarket structure
The humanoid race has entered deployment.
Price, reliability, useful hours, and learning velocity now matter more than another walking demo. China is already competing across the full price ladder.
Explore 30 humanoids02 / Intelligence library
One question per note.
One page per answer.
Filter the archive by strategic lens, or search across titles, topics, and core questions.
Strategic lens
10 / 10
Data engines · Embodied AI
Three fuels for embodied intelligence
Why no single source of robot experience is enough, and where the durable data advantage may live.
Architectures · Control
Robots that think fast and slow
A useful map of the gap between deliberate reasoning and millisecond physical control.
Scaling laws · Economics
Robotics is pre-Chinchilla
Robot learning still lacks a reliable recipe for allocating model, compute, simulation, and fleet hours.
Form factors · Humanoids
Why robots shaped like us?
The humanoid body is not a sci-fi preference. It is an interface to human spaces, tools, and training data.
Foundation models · Strategy
Generalist first, specialist later
Why task-specific robot policies may become distilled products of a shared general foundation.
Hardware · Edge compute
Embedded AI hardware, 2026
Twelve onboard compute platforms compared across AI performance, memory, power, price, and release.
Market landscape · Humanoids
The humanoid robot field, 2026
Thirty prominent humanoid platforms across five countries, mapped by commercial availability, price, release, compute, and disclosed deployment.
30 robots
05 countries
21 China-based
Learning systems · Home lab
Robots for learning at home
Twenty practical learning platforms compared by curriculum, openness, hardware, setup, price, and AI potential.
Foundation models · VLA
The VLA model zoo, 2026
Ten vision-language-action models compared by developer, release, size, weights, and how each turns perception into motor actions.
Trust · Embodied AI
The trust stack
Four layers for proving robot identity, outcomes, fleet integrity, and skill provenance. Build the model interactively.
No notes match this view. Try another lens or search term.
03 / Working method
Research built to survive a strategy conversation.
Start with the mechanism.
Separate what a source actually establishes from what the field merely repeats.
Translate technology into advantage.
Ask what changes in cost, speed, defensibility, product, or market structure.
Bring the question back to execution.
Connect the research to operating choices, product constraints, and the path from technical capability to useful work.