Samsung Patents a Self-Driving System That Uses an AI Language Model to Double-Check Its Own Routes
Samsung has filed a patent for a self-driving system that doesn't just plan a route, it then asks an AI language model to review that route for logic flaws and safety problems before the car ever moves.
What Samsung's AI route-checker actually does for self-driving cars
Imagine your car's navigation system drafts a path, then immediately asks a second AI to poke holes in it before you go anywhere. That's roughly what Samsung is describing here.
The system starts by reading camera images of the road ahead and sketching an initial driving path. It then sends that draft to a large language model (the same kind of AI behind chatbots) with two specific questions: does this path make logical sense for the driving situation, and is it actually safe? The model works through those two questions in order, revising the path twice before settling on a final route.
The idea is that a single-pass planner can miss edge cases, so building in a structured review step (first check the logic, then check the safety) should catch problems that a simple first draft might not. Whether it works better in practice than existing approaches is a question the patent doesn't answer, but the structure is clear.
… obtaining a first query command comprising a first query for a logic of trajectory generation and a second query for safety of the trajectory …
Translation: The system asks itself two questions about why it chose a path and whether that path is safe.
How the LLM runs two back-to-back safety queries on each path
The patent describes a three-stage pipeline for generating a driving trajectory (the precise path a self-driving car will follow).
Stage one: cameras collect images of the surrounding environment, and the system converts that visual data into a scene token (a compressed numerical representation of what the car sees, similar to how an AI image model encodes a photo into data it can reason about). Simultaneously, a first rough trajectory is generated from that same image data.
Stage two: the system assembles a query command containing two distinct questions for a large language model (LLM):
- A logic query: does this trajectory make sense given the driving context?
- A safety query: does this trajectory avoid hazards and meet safety requirements?
Stage three: the LLM processes those queries sequentially. The logic query produces a revised second trajectory. The safety query then processes that second trajectory to produce a final, third trajectory that the car actually follows.
The key detail is that both queries are answered using the same scene token and first trajectory as context, but the safety check explicitly builds on the output of the logic check rather than starting fresh, creating a layered review rather than two parallel but disconnected passes.
… generating a third trajectory based on the scene token, the first trajectory, and the first query command using a large language model (LLM) …
Translation: A language model builds a final driving path by reviewing its own earlier route decisions.
What this means for AI judgment calls behind the wheel
The hardest part of self-driving isn't straight roads on a sunny day, it's the edge cases that a single planning pass can't anticipate: a construction worker stepping into a lane, a car stopped halfway through an intersection. A system that explicitly separates logical correctness from safety evaluation before committing to a path addresses that problem in a structural way, not just by training a bigger model and hoping it generalizes.
For the average person, this is still a distant technology. But for automotive engineers and Samsung's own chip and sensor divisions, a patented method that routes LLM reasoning directly into trajectory planning could inform how future systems balance speed of decision-making against thoroughness of safety checking. Samsung's filing activity in autonomous driving AI has been growing, and this one sits at the intersection of the company's semiconductor, camera, and AI work.
This is the seventh Samsung filing we've tracked since July on our teams of AI models, following one on blending AI models without retraining and one on auditing answers before speaking.
Self-driving car crashes rarely happen because the car moved too slowly or missed a stop sign. They happen because the car committed to a plan and had no internal process for asking whether that plan was actually safe. That failure has cost lives, triggered congressional hearings, and wiped out companies.
Samsung's patent addresses exactly that gap by building a two-step review into the car's decision-making: first checking whether the planned route makes logical sense, then checking whether it is safe, before the car moves. The cost of skipping that review is already visible in the public record.
Whether this runs fast enough to matter on a real road is an open question the filing does not answer. But the problem it is aimed at is real, well-documented, and expensive enough that a serious attempt to solve it deserves attention on those grounds alone.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
10 drawing sheets from US 2026/0264721 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →