01 /
HUMAN INTENT
How machines understand what humans actually want.
nodes // human-intent → planning
← Intent to Real WorldEnter the map ↓
Intent to Real World / Research
RealWorldAtlas
Mapping the problems intelligence must solve before it can operate in the physical world.
Research.Systems.Machines.People.Unsolved problems.
Atlas status // active
Last updated // 2026-08-19
Research origin // Milan, Italy
We study where intelligence breaks when it meets reality.
AI has become extraordinarily good at understanding information. The physical world is different.
- It moves.
- It fails.
- It changes.
- It requires permission.
- It involves humans.
- It involves money.
- It contains uncertainty.
- Outcomes must happen.
Real World Atlas maps the research, systems, companies and unresolved problems defining this transition — from digital intelligence to physical agency.
Real World Atlas
Milan / Italy
Research status // active
// Field note 041
The physical worlddoes not returnclean API responses.
Milan · 45.4642° N, 9.1900° E
Atlas // problem map
The real world problem map.
Twelve stages sit between a human intent and a verified physical outcome. Around them, the research domains working on each stage. Filter by stage, category and research maturity — every view has its own shareable address.
Stage
Category
Maturity
View // unfiltered
https://intenttorealworld.com/atlas#mapMap // no filter applied
Select a stage on the spine — or a category and maturity level — to filter the domains, open problems and research attached to it.
Domains 18
Categories 12
Open problems 19
Dossiers 07
Published notes 00
Connected research domains
- ROBOT LEARNING
- HUMANOIDS
- AUTONOMOUS MOBILITY
- ROBOT MANIPULATION
- DEXTERITY
- HOME ROBOTICS
- HOSPITALITY ROBOTICS
- DELIVERY ROBOTICS
- EMBODIED AI
- WORLD MODELS
- SIMULATION
- VISION-LANGUAGE-ACTION MODELS
- HUMAN-ROBOT INTERACTION
- HUMAN-IN-THE-LOOP
- AGENTIC COMMERCE
- REAL-WORLD SERVICES
- PHYSICAL AI INFRASTRUCTURE
- SENSING & TELEMETRY
Atlas categories
- 01 / HUMAN INTENT
- 02 / EMBODIED INTELLIGENCE
- 03 / MANIPULATION & DEXTERITY
- 04 / WORLD MODELS
- 05 / LONG-HORIZON EXECUTION
- 06 / AUTONOMOUS MOBILITY
- 07 / HUMAN ↔ MACHINE
- 08 / REAL-WORLD ORCHESTRATION
- 09 / FAILURE & RECOVERY
- 10 / VERIFICATION
- 11 / ROBOTICS ECONOMICS
- 12 / PHYSICAL AI INFRASTRUCTURE
Open problems
- How does a machine know what a human really means?
- How much of a scene must be understood before acting is safe?
- How does a plan stay valid in an environment that keeps moving?
- Why is contact-rich manipulation still harder than language?
- How do autonomous systems handle the long tail of physical space?
- Who is an agent, legally and technically, when it acts for someone?
- Who gives an autonomous agent permission to spend money?
- How does a machine purchase a real-world service end to end?
- What are the limits of delegated authority for a physical action?
- When should execution move from machine to human?
- How do machines coordinate with humans who do not behave deterministically?
- How should robots interact with businesses that have no machine-readable interface?
- How does an autonomous system operate where APIs do not exist?
- How does physical AI recover when reality diverges from the plan?
- What does a retry mean when the first attempt changed the world?
- How does a system verify that a physical outcome occurred?
- Which cost curve makes physical execution economically inevitable?
- How do physical agents establish trust with people and businesses?
- How does an agent choose between human and machine execution?
Research dossiers
- When a human says what they want, does the machine actually understand the goal?Dossier 01 // UNVERIFIED
- What happens after step one?Dossier 02 // UNVERIFIED
- What happens when reality says no?Dossier 03 // UNVERIFIED
- Why is the human hand still so difficult to replace?Dossier 04 // UNVERIFIED
- When the driver disappears, what infrastructure remains?Dossier 05 // UNVERIFIED
- How does a machine buy something in the real world?Dossier 06 // UNVERIFIED
- Who gives AI permission to act in the real world?Dossier 07 // UNVERIFIED
Published research notes
STATUS // AWAITING VERIFIED SOURCES
Execution spine
PERCEPTION
Reading a physical scene accurately enough to act inside it.
HUMAN INTENT
Recovering what a person actually wants from ambiguous, underspecified natural language.
WORLD UNDERSTANDING
Holding a representation of an environment that keeps changing.
PLANNING
Turning a goal into an ordered sequence of physically possible steps.
IDENTITY
Who an agent is when it acts on someone else's behalf in the world.
PERMISSIONS
What an autonomous system is allowed to do, and who authorised it.
PAYMENTS
How a machine pays for a real-world service under a mandate and a limit.
ORCHESTRATION
Coordinating humans, providers, APIs and machines toward one outcome.
PHYSICAL EXECUTION
The moment intelligence touches matter: motion, contact, transport.
FAILURE RECOVERY
What happens when reality diverges from the plan mid-execution.
VERIFICATION
Proving the intended physical outcome actually occurred.
LEARNING
Feeding real-world outcomes back into the system that caused them.
Connected research domains
ROBOT LEARNING
Policies learned from data, demonstration and interaction.
HUMANOIDS
General-purpose embodiment in human environments.
AUTONOMOUS MOBILITY
Moving people, objects and machines through physical space.
ROBOT MANIPULATION
Grasping, placing and acting on objects.
DEXTERITY
Fine contact-rich control; hands as the hardest interface.
HOME ROBOTICS
Unstructured private environments with no schema.
HOSPITALITY ROBOTICS
Service execution where guests, staff and machines share a floor.
DELIVERY ROBOTICS
Last-metre handover between machines and people.
EMBODIED AI
Intelligence that reasons from inside a body in an environment.
WORLD MODELS
Learned predictive representations of physical dynamics.
SIMULATION
Synthetic environments, and the gap back to reality.
VISION-LANGUAGE-ACTION MODELS
Language and vision mapped directly onto action.
HUMAN-ROBOT INTERACTION
Legibility, trust and handover between people and machines.
HUMAN-IN-THE-LOOP
When execution should escalate from machine to person.
AGENTIC COMMERCE
Identity, mandate and payment for autonomous buyers.
REAL-WORLD SERVICES
Providers, availability and businesses without machine interfaces.
PHYSICAL AI INFRASTRUCTURE
The layer underneath intelligence that operates in the world.
SENSING & TELEMETRY
The instruments that report physical state back.
Motion is not the hard part.
Knowing it happened is.
Physical execution plate
Robotic manipulation · illustrative reference, third-party hardware
Atlas // taxonomy
Twelve domains under observation.
02 /
EMBODIED INTELLIGENCE
How intelligence perceives and reasons inside physical environments.
nodes // perception → world-understanding
03 /
MANIPULATION & DEXTERITY
How machines physically interact with objects.
nodes // physical-execution
04 /
WORLD MODELS
How intelligent systems represent and predict physical environments.
nodes // world-understanding → planning
05 /
LONG-HORIZON EXECUTION
How machines complete tasks spanning many steps, changing state and time.
nodes // planning → orchestration → failure-recovery
06 /
AUTONOMOUS MOBILITY
How machines move humans, objects and themselves through physical space.
nodes // physical-execution → perception
07 /
HUMAN ↔ MACHINE
How humans and intelligent machines coordinate.
nodes // human-intent → permissions → verification
08 /
REAL-WORLD ORCHESTRATION
How intelligence coordinates external systems, providers, APIs, humans and machines.
nodes // orchestration → payments
09 /
FAILURE & RECOVERY
What happens when physical execution does not go according to plan.
nodes // failure-recovery
10 /
VERIFICATION
How a system knows the intended physical outcome actually happened.
nodes // verification
11 /
ROBOTICS ECONOMICS
Cost curves, deployment economics, labour substitution and complementarity, utilisation, scaling.
nodes // orchestration → learning
12 /
PHYSICAL AI INFRASTRUCTURE
The infrastructure required underneath intelligence operating in the physical world.
nodes // identity → permissions → payments → orchestration → verification
Atlas // research dossiers
The first seven frontiers.
Seven questions, opened in order. Each dossier holds its research areas, its Atlas nodes, its open problems and our thesis — and stays empty of findings until a primary source has been read and attributed.
- Dossier 01HUMAN INTENTWhen a human says what they want, does the machine actually understand the goal?nodes // human-intent → planning → physical-executionStatus // readingopen →
- Dossier 02LONG-HORIZON EXECUTIONWhat happens after step one?nodes // planning → world-understanding → orchestration → physical-executionStatus // readingopen →
- Dossier 03FAILURE & RECOVERYWhat happens when reality says no?nodes // failure-recovery → world-understanding → verificationStatus // readingopen →
- Dossier 04HUMANOID DEXTERITYWhy is the human hand still so difficult to replace?nodes // physical-execution → perception → learningStatus // readingopen →
- Dossier 05AUTONOMOUS MOBILITYWhen the driver disappears, what infrastructure remains?nodes // physical-execution → identity → payments → verificationStatus // readingopen →
- Dossier 06MACHINE COMMERCEHow does a machine buy something in the real world?nodes // identity → permissions → payments → orchestration → verificationStatus // readingopen →
- Dossier 07flagshipPERMISSION TO ACTWho gives AI permission to act in the real world?nodes // identity → permissions → payments → orchestration → verificationStatus // readingopen →
Atlas // research chain
- VERIFICATION
The dossiers are not isolated articles. Each one inherits the failure of the stage before it.
// System observation 052
Reality isthe hardest API.
Atlas // latest signals
Latest signals.
Signal queue // status: reading
No research note is published until its primary sources have been read and attributed. The Atlas structure, taxonomy, map and open-problem layer are live; notes enter the index one verified source at a time.
Source first. Claim second. Interpretation third.
Source hierarchy
- LEVEL 01 — Peer-reviewed research, arXiv from credible groups, university labs, conference proceedings
- LEVEL 02 — Official technical research from serious robotics / AI organisations
- LEVEL 03 — Technical interviews with researchers, engineers and founders
- LEVEL 04 — High-quality analysis from respected research institutions or investors
Institutions appear as research sources only. Never as partners.
Research maturity system
- 01 / THEORY
- 02 / SIMULATION
- 03 / LAB
- 04 / CONTROLLED REAL WORLD
- 05 / FIELD DEPLOYMENT
- 06 / COMMERCIAL
- 07 / SCALE
Insufficient evidence // status: unverified
Atlas // unsolved
What still doesn't work.
Most technology pages celebrate progress. The Atlas maps failure. Each item below is an open problem, not a product claim, and none of them are presented as solved.
Open problem
INTENT
How does a machine know what a human really means?
human-intent
Open problem
PERCEPTION
How much of a scene must be understood before acting is safe?
perception
Open problem
PLANNING
How does a plan stay valid in an environment that keeps moving?
planning → world-understanding
Open problem
DEXTERITY
Why is contact-rich manipulation still harder than language?
physical-execution
Open problem
NAVIGATION
How do autonomous systems handle the long tail of physical space?
physical-execution → perception
Open problem
IDENTITY
Who is an agent, legally and technically, when it acts for someone?
identity
Open problem
PAYMENTS
Who gives an autonomous agent permission to spend money?
payments → permissions
Open problem
PAYMENTS
How does a machine purchase a real-world service end to end?
payments → orchestration
Open problem
PERMISSIONS
What are the limits of delegated authority for a physical action?
permissions
// Open problem 071
Who gives a machinepermission to act?
Research note // 001
Who gives AI permission to act in the real world?
Identity, delegated authority, agent payments, permissions and verified real-world execution.
Read the research note →Research note // 002
If atoms become bits, what becomes the operating system?
Physical compute, real-world scheduling and the missing layer between human intent and the physical world.
Read the research note →Research note // 003
Reality doesn't throw exceptions.
Failure recovery, replanning, closed-loop execution, executor substitution and verified outcomes in Physical AI.
Read the research note →Atlas // thesis
Intelligence is accelerating.Reality is not.
Model capability↑↑↑
Real-world execution↑
The execution gap
- Identity
- Payments
- Permissions
- Providers
- Physical access
- Human coordination
- Failure recovery
- Verification
We do not claim these layers are solved. We are mapping them, and building inside them.
This is where we are looking.
Autonomy moves matter.
Infrastructure decides whether it counts.
Autonomous mobility plate
Autonomous mobility · illustrative reference, third-party hardware
// Reality 084
A robot can failwithout throwingan exception.
We don't study robotics to predict robots. We study it to understand what the world will require.
Every breakthrough changes what machines can do.
Every new capability creates new infrastructure requirements.
- Perception created data infrastructure.
- Reasoning created agent infrastructure.
- Physical action will create real-world execution infrastructure.
We are mapping that transition.
// Thesis 091
The last mileof intelligenceis reality.
Atlas // origin
Why we built this.
We operate in the real world.
That means we care less about what AI can demonstrate and more about what intelligence can reliably cause to happen.
Real World Atlas is our attempt to understand the systems, breakthroughs and unsolved problems shaping the transition from digital intelligence to physical agency.
We study the frontier because we are building toward it.
Built quietly in Milan
Intent to Real World
/ Research
Building something that touches the real world?
- Robotics.
- Physical AI.
- Autonomous systems.
- Embodied intelligence.
- Real-world infrastructure.
We should probably talk.
Continue // technical record
// Milan, Italy
// Real World Atlas
// Observation continues