Patents
These patents came from problems we worked on at Phantom: keeping the connection usable, helping operators, and handling failures. Fourteen inventions were issued as twenty-two US patents between 2021 and 2023. I have grouped them by the problems they address.
The vehicle encodes its camera feed several times over and sends one copy down each cellular connection it has. When a carrier degrades, the operator's video holds, because the same frames are already arriving by another route.
This is the server half of the same link. It decides how to spread traffic across whatever connections are up, trading bandwidth against delay, so the feed stays usable over networks nobody in the loop controls.
The operator sets a path in a simulation, and the vehicle follows it using its own control system. That reduces reliance on continuous manual input.
Software on the vehicle rewrites incoming commands using sensor data the remote side cannot see, such as load weight or wheel slip. A forklift and a yard truck then feel close enough to the same machine that an operator can move between them without relearning the controls.
The system builds a depth map of what surrounds the vehicle, works out the fastest it could travel and still be stopped in time by a human who is not standing there, and caps it at that. How long the round trip to the operator takes is one of the inputs, so the cap tightens as the network gets worse.
The encoder finds background objects an operator does not need in order to drive, deletes those pixels, and sends a short description in their place. The receiving end renders stand-ins so the scene still reads as a whole street.
Vehicles get matched to operators by an optimizing function instead of a queue. It can also put several operators on one vehicle at once, so the vehicle takes whichever command stream arrives first, or a proxy builds a consensus command out of where the operators disagree.
A vehicle, or a person standing next to one, can call for a human. Scanning a QR code or an RFID tag on the machine opens the session, and the system decides how urgent the request is before it pulls an operator off other work.
Available operators get scored against the specific request, on weather, vehicle class, and their own history, and the best match is assigned. One embodiment runs the whole thing as an exchange where operators bid for the jobs they want.
Every drive records how the network actually behaved at each point on the map. A vehicle reads its own position against that history and adjusts its congestion window and bit rate before the dead spot rather than after it.
Routing by predicted signal quality rather than by distance, so a remotely driven vehicle takes the road where the link will hold instead of the road that is shorter. The operator can also see the prediction as a map overlay and pick for themselves.
A policy server holds the rules about where a given vehicle may not go, updatable over an API rather than by reflashing anything. Enforcement lands on the onboard computer, on the teleoperation console, or on both.
Put enough remotely driven vehicles in one neighborhood and they start competing for the same cell capacity. This allocates it with game-theoretic modelling, rewarding the vehicles that hold to the sharing policy and penalizing the ones that take more than their share.
Teleoperated and autonomous vehicles working inside a shipping yard, with drones holding position overhead to relay signal past the container stacks that block it. The operator works from an augmented view of the yard instead of raw camera feeds.
Filed as Shay Magzimof, which is the spelling on every US application. Each of these has co-inventors and the filings name them; none of it was mine alone. Serve Robotics holds the portfolio after buying Phantom Auto's assets in 2025. Two earlier grants lapsed after maintenance fees went unpaid. They remain part of the work we did.