The hardest part of building a self-driving car is not the sunny-day driving; it is the moment the system runs out of confidence — a delivery van double-parked across a lane, a lane line that vanishes into construction, weather that degrades the sensors. Two industry answers exist for that moment, and they are not the same product. In remote assistance, a human operator sitting in a control room answers a narrow question (“is it safe to go around that parked vehicle?”) while the car keeps driving itself. In remote driving, the operator takes over steering and pedals over a network link. A Toyota grant that issued in the June 30, 2026 USPTO grant drop, US12669818B2 (“Remote control request system, remote control request method, and nontransitory storage medium”), is directed at the deciding question underneath both: which of those two kinds of help should the car ask for, and when.

The disclosed method makes that call from the vehicle’s autonomous driving domain — the engineering term of art (operational design domain, or ODD) for the conditions under which the car is allowed to drive itself. The granted independent claim encodes that domain as an explicit gate. The car is permitted to keep driving autonomously only when a set of conditions all hold: the road’s speed limit is below a threshold, map information is available for the route, and there is no rain currently falling or forecast in the area for the relevant window. If the car is inside that domain and hits an ambiguity — the claim’s worked example is another vehicle parked on the shoulder — it requests remote assistance and keeps driving. If it is outside the domain, it instead requests remote control driving and hands the operation to the operator.

A remote control request method has requesting, by a computer, a remote operator to perform remote control on an autonomous driving vehicle when the autonomous driving vehicle currently has or is expected to have difficulty in continuing autonomous driving, requesting remote assistance in which the remote operator makes at least a part of determination for the autonomous driving inside an autonomous driving domain, and requesting, outside the autonomous driving domain, remote driving in which the remote operator performs at least one of a steering operation and an acceleration or deceleration operation of the autonomous driving vehicle.— Remote control request system, remote control request method, and nontransitory storage medium, US12669818B2

Why split “assistance” from “driving” at all?

The split is a response to the physics of a network link. Remote driving puts a human in a closed control loop across cellular latency and variable bandwidth; every steering input is delayed by the round trip, which is tolerable at low speed but not at highway speed. Remote assistance keeps the fast control loop onboard the car and asks the operator only for a discrete judgment, which is far more forgiving of lag. Reading the claim as an engineer, the three gating conditions map almost one-to-one onto the factors that make remote driving risky: a low speed limit bounds how fast a delayed input can go wrong, map availability guarantees the car has priors to fall back on, and the no-rain condition protects the perception stack the whole scheme depends on. The method, in other words, does not treat the human operator as a single fallback; it grades the fallback and picks the lighter-touch version whenever the domain conditions allow it. One point is load-bearing and worth stating plainly: this is an issued grant (kind code B2), not a pending application, so the domain-gated switch above is claim language the office has allowed, classified under the autonomous-control classes G05D 1/0061 and G05D 1/2279.

The perception and reasoning the decision rides on

A request-for-help decision is only as good as the car’s read of the scene that triggered it, and the same grant drop carries a cluster of Toyota-group perception grants that feed exactly that read. US12670795B2 (“Monitoring device”) describes watching two moving objects around the host vehicle and deciding to warn them when an obstacle interrupts each one’s detection of the other — the blind-corner case where two road users are on a collision course neither can see. US12670793B2 (“Alarm device and alarm method”) tackles the arbitration problem that follows: when several alarm targets appear at once, it assigns priorities and suppresses a lower-priority warning rather than firing everything. US12669336B2 reads road geometry, detecting that a branching road has appeared by tracking the vanishing point where two lane lines intersect — the kind of structural change that can push a car toward the edge of its map coverage.

Below that sits the object-recognition layer. US12670738B2 (“Image processing device of person detection system,” assigned to Toyota Industries) switches between an upper-body and a whole-body detector depending on how far a pedestrian stands from the road surface, a pragmatic fix for the fact that a distant person fills too few pixels for a full-body match. Two grants from Toyota Research Institute round out the reasoning side: US12670211B2 (“Determining query complexity in video question answering”) builds an abstract syntax tree from a natural-language query about a video and scores how hard it is to answer, and the clustering method in US12670187B2 groups high-dimensional vector data by density. Even US12671792B2 (“Space coupling system”), which couples a real space to a virtual one around a moving person or screen, reads as infrastructure for the operator’s side of a teleoperation link. Taken together, the cluster describes a company engineering the seam between an automated driver and a human one — and treating ‘when to ask for help, and what kind’ as a first-class control problem rather than an afterthought.