Skip to content
Multigrid

Robotics & embodied AI

Why putting a model in a body is a different problem from putting one behind an API — the control deadline, the missing data, the simulation gap and the fact that failure is physical.

A language model has no deadline. If it takes two seconds instead of one, a person waits a second longer and nothing else in the world changes. A robot has a deadline set by physics: an arm carrying a kilogram at half a metre per second travels a centimetre every twenty milliseconds whether or not anything has decided what it should do next, and a controller that misses its period is not slow, it is unstable. Almost everything that makes embodied AI hard follows from that sentence, plus the fact that nobody could scrape the training data.

These pages are about the mechanics. Where a rate, a latency or a data volume decides an architecture, the arithmetic is done in front of you from assumptions you can replace. Where a research result is the reason something works, it is named. What you will not find is a report on what any particular robot can do this month — that is the part of robotics writing that is already abundant, and it is wrong within a quarter.

Embodied AI: Why the Physical World Is Harder

Why a model that writes a correct paragraph and a robot that picks up a mug are not the same problem at two difficulties, told through the deadline, the data and the cost of being wrong.

5 min read

Vision-Language-Action Models: How a Transformer Ends Up Producing Torque

The path from a vision-language backbone to a command an actuator can accept, including the two ways actions get represented and why the output is a chunk rather than a step.

5 min read

The Robotics Data Problem: Why There Is No Internet of Torques

Why manipulation has no scrapeable corpus, what the orders of magnitude actually are against language pretraining, and what each of the four available sources costs.

5 min read

Sim-to-Real Transfer: What Crosses the Gap and What Does Not

The sim-to-real gap named part by part — contact, friction, actuator dynamics, latency and sensor noise — with the techniques that close each and the cost each technique charges.

6 min read

Imitation Learning and Teleoperation: Copying a Human at a Cost That Scales

Behaviour cloning, the compounding-error result that explains why it fails, and what each style of teleoperation rig silently omits from the data it produces.

5 min read

Reinforcement Learning for Control: Sample Efficiency Is the Binding Constraint

Why RL is attractive for control and why almost nobody runs it on hardware, done as an arithmetic problem in robot-hours, plus the four ways the field routes around it.

5 min read

Humanoid Robots: The Case For and Against the Form Factor

The engineering arguments for and against a human-shaped machine, including the reliability and energy arithmetic that a form-factor decision actually turns on.

4 min read

Manipulation: Why Hands Are Still Hard

What happens mechanically in the last centimetre before contact, why vision stops helping exactly there, and why grasping has resisted learning in a way navigation did not.

5 min read

Autonomous Vehicles: How to Read a Claim About Where the Field Is

What the SAE levels and the operational design domain actually commit a system to, why disengagement rates do not compare across programmes, and the statistical reason safety claims are so hard to establish.

5 min read

Drones and Aerial Autonomy: When the Compute Has to Fly With You

Why aerial autonomy is shaped by two hard numbers — the attitude loop rate and the power cost of carrying a processor — and what those force the algorithms to look like.

4 min read

Warehouse Robots: What Changing the Building Bought

How goods-to-person fulfilment worked by re-engineering the environment around the machine, which problems that deleted outright, and which one it left standing.

5 min read

Latency in Robotics: When 100 ms Is Too Slow

What a control deadline is, how much stability a delay costs in phase margin, where the milliseconds go in a real loop, and what fills the twenty cycles while a model is thinking.

6 min read

Edge Compute for Robots: What Fits On Board and What Must Go Remote

The watt, gram and memory-bandwidth budget that decides what a robot can run locally, with the bandwidth calculation that caps generation rate, and a rule for splitting the stack.

5 min read

Safety Standards for Physical AI: Where Certification Meets a Stochastic Policy

How functional safety works, why a learned policy cannot be certified the way a controller can, and the architecture that lets an uncertifiable component drive a certified machine.

6 min read

Foundation Models for Robotics: Stating the Bet So It Can Be Settled

The transfer-learning bet behind generalist robot policies, the lineage of architectures that made it, the cross-embodiment problem in the way, and what would confirm or refute it.

6 min read

Robotics Simulators: The Axes They Actually Differ On

A comparison built on the architectural choices that decide what a simulator is good for — contact formulation, parallelism, rendering, determinism — rather than on a feature list that expires.

6 min read

Benchmarks in Robotics: Why the Numbers Are So Hard to Compare

Why robot results resist comparison — the hardware is part of the benchmark, trials are expensive, and twenty of them barely distinguish anything — plus what a usable report must contain.

5 min read

What a Deployed Robot Costs: The Payback Formula With Every Term Named

A payback model for an automation cell in which every cost term is enumerated and every figure is a reader-supplied input, plus the three terms that usually decide the answer.

5 min read

Teleoperation as a Bridge Technology: The Ratio That Decides It

Why remote human operation is an operating model rather than an admission of failure, the staffing ratio that makes or breaks it, and the delay physics that limits what an operator can do.

5 min read

Will Robotics Have Its GPT Moment? Four Conditions to Watch

What actually produced the language-model discontinuity, which of those conditions robotics currently lacks, and the observable signals that would indicate the gap closing.

5 min read

Other topics

Robotics & embodied AI · Multigrid