ドメインギャップを超えて：AntigravityとGeminiで構築したAIレースコーチBridging the Domain Gap: AI Race Coach built with Antigravity and Gemini

Antigravity has published a case study describing how it built an AI-powered race coach on top of Google's Gemini models, offering a concrete look at the challenges of adapting a general-purpose large language model to a highly specialized field. The project matters because it illustrates a problem that affects almost every serious attempt to deploy LLMs in professional settings: the so-called domain gap between what a model learns during broad pretraining and what an expert user actually needs.

Motorsport coaching is a demanding test case. A human coach studies lap times, telemetry, and onboard video to tell a driver where to brake later, how to carry more speed through a corner, or how to adjust throttle on corner exit. Much of this knowledge is tacit, context-dependent, and expressed in a vocabulary that a general model may understand only partially. A model trained on the open web can discuss racing in broad terms, but it does not natively reason about a specific car, circuit, or session with the numerical precision a driver expects.

The domain gap, as the case study frames it, appears in several forms. There is a vocabulary gap, where terms such as trail braking, apex, or understeer carry precise operational meaning. There is a data gap, because raw telemetry arrives as long numerical time series rather than natural language. And there is a reasoning gap, since useful coaching requires connecting patterns in that data to concrete advice grounded in physics and driving technique. Each of these has to be addressed before the output feels credible to an experienced driver.

To bridge these gaps, the team appears to combine several techniques rather than relying on any single one. Retrieval-augmented generation, commonly abbreviated as RAG, lets the system ground its responses in curated coaching material and reference data, reducing the tendency to produce plausible-sounding but incorrect guidance. Structured preprocessing converts telemetry into summaries, engineered features, or annotated segments that the model can interpret more reliably than a raw stream of numbers. Carefully written system instructions and prompts encode the persona, priorities, and safety boundaries of an expert coach so the assistant stays focused and consistent.

Gemini's multimodal capabilities are central to this approach. Because the model family can process text alongside other inputs and handle long contexts, a coaching system can bring together varied signals in a single reasoning step rather than stitching outputs from separate tools. That is well suited to motorsport, where meaning emerges from the relationship between what a driver did, what the data recorded, and what an ideal reference lap looks like. The case study is likely of interest less for the specific racing use case and more as a template for any team facing similar constraints.

It helps to place this work against adjacent tools and concepts. Fine-tuning, where a base model is further trained on domain examples, is an alternative or complement to RAG; the two are often weighed against each other on cost, data availability, and how frequently the underlying knowledge changes. Function calling and tool use let a model query databases or run calculations rather than guessing at numbers, which is valuable when precision matters. Evaluation is another prerequisite that is easy to underestimate, since judging whether coaching advice is genuinely good requires domain experts, not just automated metrics.

The broader context is a steady industry shift from generic chatbots toward grounded, task-specific assistants. Motorsport already relies heavily on data analytics, with professional teams employing simulation, telemetry platforms, and performance engineers, so layering an LLM interface onto that pipeline is a natural extension rather than a departure. Similar patterns are emerging across medicine, law, finance, and engineering, where practitioners want systems that respect specialized terminology and cite verifiable sources.

Readers should treat any case study as a curated account that highlights successes, and independent details on accuracy and real-world driver outcomes are limited here. Even so, the write-up is a useful, practical illustration of how vocabulary, data representation, retrieval, and prompting can be assembled to narrow the domain gap when applying models like Gemini to expert work.