Guides to physical AI

Why these exist

This venture has no product to describe, and writing product pages for software that does not exist would be dishonest. What it can do is explain the problem clearly. These guides set out how robot software is actually built, what the current approaches genuinely achieve, and where each one breaks — including the parts that are usually compressed into a sentence about ongoing work.

They are written for people deciding whether any of this is real: engineers assessing the approaches, manufacturers weighing where the software layer should sit, and operators trying to work out whether a robot could do a particular job. Where the honest answer is that something does not work yet, the guides say so.

The development status page records what exists in this venture, which is separate from what exists in the field.

How this site is written records how claims here are sourced, and what has and has not been reviewed by a qualified professional. The short answer is that nothing has.

Finding your way around

There are three ways through this. Start here is the short route if you have a decision in front of you. The five groups below sort every guide by what you are trying to do. And the full alphabetical list is still at the bottom if you would rather just scan it.

Start here

Describe one capabilityWhat a capability does, what it needs, where its competence ends, and who fixes it. Template included.Scope an integrationThe facts anyone will ask for before they can give you a number. Questionnaire included.Work out the rightsWhat is being licensed, what the dependencies oblige, who owns the data. Rights inventory included.Judge a pilotWhat to agree before it starts, so the number it produces means something. Scorecard included.

Understand the stack8

What the pieces are and how they fit. Start here if you are working out what robot software consists of.

What is physical AI? — Artificial intelligence that acts in the physical world, and why that makes it a fundamentally harder engineering problem than software-only AI.
What is embodied AI? — Where the term embodied AI comes from, what embodiment actually changes about a learning problem, how it differs from physical AI, and what the field has and has not demonstrated.
The humanoid robot software stack — What sits between a humanoid robot and useful work: the layers, what each is responsible for, and where the hard boundaries are.
ROS 2 and robotics middleware — What middleware does in a robot, how ROS 2 works, what it solved, and the problems it was never designed to solve.
The robot perception stack — How a robot turns sensor data into something it can act on, which sensors fail in which ways, and why calibration causes more trouble than algorithms.
Robot motion planning — How a robot decides how to move: configuration space, the main algorithm families, and why planning is rarely the thing that fails.
SLAM and robot localization — The difference between mapping, localization and SLAM, the sensor families used for each, the named systems worth knowing, and the failure modes that decide whether a deployment survives contact with a real building.
Onboard compute for robots — Why robots compute locally, how to size power, thermal and latency budgets, what the Jetson Orin and Thor families actually offer, and the separation between the safety controller and everything else.

Specify a capability8

Defining one piece of work precisely enough that two organisations mean the same thing by it.

Robot capability specification template — Define a robot capability with a reusable template covering interfaces, hardware assumptions, operating limits, evaluation evidence and ownership.
Robot software integration requirements checklist — Collect robot models, interfaces, sensor details, dependencies and test goals before scoping a robot software integration.
Robot grippers and end effectors — Parallel jaw, vacuum, multi-finger, soft and bespoke end effectors compared, what tactile and force sensing actually adds, and why the gripper decision constrains the whole project.
Why robot grasping is hard — Picking something up looks trivial and is not. What makes grasping difficult, why benchmark numbers mislead, and what actually works.
Robot foundation models — The attempt to build general-purpose models for robot control: what the idea is, what the evidence shows, and what it has not solved.
Vision-language-action models — How VLA models work, why they became the dominant approach to learned robot control, and what they are still bad at.
Imitation learning for robots — Teaching robots from demonstrations: why it became the default, how the main methods differ, and the failure mode built into the approach.
Reinforcement learning in robotics — Learning from reward rather than demonstration: where it genuinely wins, why it is mostly done in simulation, and what makes it hard to deploy.

Evaluate a workflow5

Finding out whether something works before committing to it, and what the answer costs.

Robot pilot acceptance criteria and scorecard — Plan a robot pilot with defined tasks, baselines, intervention rules, evidence requirements and a practical acceptance scorecard.
Robot simulation software — The robot simulator landscape as it actually stands: what Gazebo, Isaac Sim, MuJoCo, Drake and the rest are each good at, how to choose, and the costs nobody includes in the estimate.
Sim-to-real transfer — Why robotics depends on simulation, exactly where simulation lies, and the techniques used to cross the gap.
Teleoperation and robot data collection — Where robot training data actually comes from, what the collection rigs trade off, and why the bottleneck has not gone away.
Robotics ROI and deployment cost — What a robot deployment actually costs beyond the robot, how to build a return calculation that survives contact with reality, and the costs that get left out.
Humanoid robot manufacturers — How to evaluate a humanoid platform rather than compare specifications, and an honest view of where commercial deployment actually stands.

Deploy and operate2

Running it on real machines, on a real site, once the interesting part is over.

Robot software deployment and rollback checklist — Prepare a robot software release with version records, staged validation, rollback planning, operator handover and post-release checks.
Robot fleet management — What changes when one robot becomes fifty: traffic, task allocation, versioning, observability, and the problems that only appear at scale.
Warehouse automation software — The software layers inside an automated warehouse, what each one owns, and why integration between them is where most projects actually fail.

Safety and governance5

Standards, assessment and rights. The four safety pages here are limited reference overviews and have not been reviewed by a qualified robotics professional; the licensing checklist has not been reviewed by a lawyer.

Robot safety standards — How ISO 10218, ISO/TS 15066 and R15.06 relate to each other, what the 2025 revision changed, and what each one is actually for.
Robot risk assessment — The process that determines whether a robot application is safe, why it comes before the equipment choice, and how it is usually done badly.
Functional safety for robots — What "safety-rated" actually means, where performance levels and safety integrity levels come from, and why a software stop is not a safety function.
ISO/TS 15066 and collaborative robot operation — What ISO/TS 15066 is, the four modes of collaborative operation it describes, why there is no such thing as a collaborative robot, and where the specification sits after the 2025 revision of ISO 10218.
Robot software licensing checklist — Use a practical checklist to clarify robot software rights, dependencies, updates, data use and support before agreeing to a licence.

All 30 guides, alphabetically