When AI moves in the world.

Physical AI is different from a chatbot. If a model steers, flies, lifts, maps, watches or acts, the trust question includes safety, liability, sensors, cybersecurity and who is allowed to deploy it.

A dark Australian autonomy test hangar with sensor cars, drones, robots and a glowing map of Australia

Do not judge physical AI by demo videos.

A polished clip can hide remote control, careful editing, safety drivers, private test tracks, hand-picked routes or lab-only conditions. For real trust, look for where the system is allowed to operate, who supervises it, what happens when it fails and who is liable.

Autonomous vehicles

Robotaxis, self-driving trucks, assisted driving and off-road autonomy. Ask whether it is fully driverless, supervised, geofenced or still a driver-assistance system.

Robots and drones

Humanoids, warehouse robots, mining robots, inspection drones and defence-adjacent systems. Ask what they can do without a human and what sensors they collect.

Physical AI models

World models, simulation systems and robotics foundation models that help machines predict movement, space and cause-and-effect before acting.

The big physical AI shelves.

This page is a starting map, not a certification list. Treat every deployment claim as dated.

Robotaxis and road autonomy

Waymo, Tesla, Zoox, Baidu Apollo, Pony.ai, WeRide
High stakes

Waymo is the clearest public robotaxi reference. Tesla's Full Self-Driving is still framed as supervised. The key beginner distinction is fully autonomous service versus driver-assistance that still needs human attention.

Trust watchRegulatory approvals, crash reporting, geofencing, safety driver rules, insurance and handover failures.
Good signsTransparent service areas, public safety cases, incident reporting and clear wording about human supervision.

Humanoid and industrial robots

Figure AI, Tesla Optimus, Boston Dynamics, Agility, Apptronik, Unitree
Watch

Humanoid robots make big claims because they look familiar. The trust test is boring but important: what job can they repeat safely, with what failure mode, in what environment, under whose supervision?

Trust watchTeleoperation disclosure, workplace injury risk, labour impact, surveillance sensors and maintenance responsibility.
Good signsClear task boundaries, safety certification, real customer deployments and visible human override paths.

Simulation and robotics foundation models

NVIDIA Cosmos and Isaac, Google DeepMind robotics, Toyota Research Institute
Infrastructure

These systems help robots learn in simulation before deployment. That can reduce risk, but it also centralises power in model, chip and simulation platforms.

Trust watchSim-to-real gaps, synthetic data limits, defence use, model access and benchmark honesty.
Good signsOpen model releases, safety research, reproducible benchmarks and careful deployment claims.

Australian autonomy builders

Advanced Navigation, Applied EV, Emesent, DroneShield, CSIRO Data61
Local watch

Australia has serious autonomy work in navigation, mining, drones, mapping, defence and research. For locals, the question is not only "Is it clever?" but "Where is it deployed and under whose rules?"

Trust watchDefence links, remote-area safety, workplace use, surveillance, export controls and procurement transparency.
Good signsLocal capability, mining safety, emergency mapping, infrastructure inspection and sovereign skills.

Questions before trusting a moving machine.

Use these before believing a deployment claim, investing in a company, approving a workplace rollout or sharing a video as proof.

1 Operational design domain

Where is it allowed to operate: city, mine, warehouse, road, hospital, school, home, factory or battlefield?

2 Human supervision

Is a person driving, supervising, teleoperating, approving actions or only watching after the fact?

3 Sensors and privacy

What cameras, microphones, lidar, radar, location data or biometric signals are collected and stored?

4 Cybersecurity

Can it be hacked, spoofed, jammed or updated remotely, and who controls emergency shutdown?

5 Liability

If it hurts someone or damages property, who pays: owner, operator, software maker, insurer or government?

6 Public evidence

Look for public trials, regulator notices, incident reports and independent testing, not only polished launch videos.

Assisted is not autonomous.

For cars, drones and robots, wording matters. "AI-powered", "autopilot", "supervised", "remote operation", "driverless", "robotaxi" and "fully autonomous" are different claims. The safest habit is to ask what a human must still do when the system gets confused.