Field notes
What we learned in San Francisco
We went to San Francisco to meet AI startups and investors, and to see how far physical AI has come. These are our notes, and the product ideas that came out of them.
In short
- SF leads in models. Applying them to factories, robots and PLCs is where teams get stuck.
- The three gaps: PLC skills, safety of probabilistic outputs, and the cost of testing on hardware.
- Our answer is three offers that can be adopted separately: Edge Box, PAI-Ground and PAI-Field.
01Riding in a car that measures the world
Riding a Waymo through the city, the striking part was not the driving. It was that the car never guesses distance. It measures it with LiDAR, many times per second, and behaves conservatively when it is unsure.
Most factory AI we see does the opposite: a 2D camera estimates depth, a language model proposes an action, and the action goes straight to the machine. That gap stayed with us for the rest of the trip.
02Where physical AI gets stuck
PLC skills
Plenty of AI engineers, very few who can write ladder logic or speak Modbus and EtherNet/IP.
Safety
Camera-based vision drifts with lighting and shadows. A probabilistic output that is wrong once can break equipment or hurt someone.
Cost of testing
Building and debugging hardware rigs locally is expensive, so test cycles stay slow.
03What we already had
Our teams in Japan and Thailand work as Forward Deployed Engineers for manufacturers, connecting PLCs, IoT sensors and LLMs on real equipment. In Japan we also run PAI-Ground, a physical test field for robotics and physical AI developers.
04Three ways in
Teams are at different stages, so we split the offer into three parts. Each solves one gap on its own, and together they form a full pipeline.
Perceive
PAI-Field
3D LiDAR point clouds for measured, not estimated, distance.
Decide
Typed guardrails
Commands with a fixed shape and a confidence score, behind a hard interlock.
Execute
Nameless Edge Box
PLC control through a REST or WebSocket API.
Test
PAI-Ground
Remote testing on real hardware in our lab in Japan.
05Why “Zero Accidents”
Factory automation is usually sold as headcount reduction. We would rather measure ourselves by something else: no one gets hurt when AI and machines share a floor with people. That is the standard we want every product above to meet.
Next
We're looking for AI teams that want to put a model on real hardware, and for investors who care about physical AI safety.