PitPerfect

Predict. Pit. Perform: AI that knows the optimal move before the tires do.

Comment

PitPerfect: The AI Co-Strategist

PitPerfect is an AI-driven co-strategist designed to master one of racing's most complex variables: tire degradation. By intelligently fusing lap telemetry, environmental data, and real-time visual analysis from onboard footage, PitPerfect provides the critical foresight needed to select the optimal pit window. It’s about transforming high-stakes gambles into data-backed, winning decisions.

Inspiration

Every F1 fan has seen brilliant drives crumble because of poor pit-stop timing — Ferrari’s 2022–23 seasons are a classic example, where strategy misjudgments cost podiums. These moments don’t come from a lack of expertise, but from the sheer complexity of predicting tire behavior. Degradation depends on countless variables — track temperature, compound type, driving style, and even minor setup changes. In the chaos of a race, strategists make high-stakes decisions with incomplete information. A single lap too early or late can change everything. That’s what inspired us to build PitPerfect — an AI co-strategist that reads between the lines of telemetry, weather data, and onboard video to understand tire wear as it happens.

Our Solution: An AI That Sees the Wear

PitPerfect is an AI-driven race strategy assistant that deciphers the complex language of tire degradation. It ingests and synthesizes dozens of variables—from track temperature and driver aggression to tire compound and car setup. Our system employs a sophisticated multimodal AI, learning critical patterns from both numerical telemetry data and, crucially, visual indicators from the car's onboard camera. It's an AI that doesn't just calculate; it sees the wear as it happens, providing a clear and dynamic tire-health index.

How It Works: The PitPerfect Pipeline

Our methodology is broken down into a high-performance pipeline:

  1. Data Collection & Preparation \ We begin by aggregating comprehensive data streams. This includes historical F1 telemetry (lap times, tire pressure, temperature) via the Ergast API, augmented with environmental data like track conditions and weather. Simultaneously, we extract and process frames from onboard video to build our visual analysis dataset. The final step is a rigorous cleaning, normalization, and time-synchronization process to create a unified foundation for our models.

  2. Predictive Model Development \ Our system's core is a dual-model intelligence layer:

    • Telemetry Model: A time-series regression network (such as an LSTM or GRU) is trained to predict degradation patterns based on sensor and lap data.
    • Visual Detection Model: A pretrained Convolutional Neural Network (e.g., ResNet, MobileNet) analyzes video frames to identify physical signs of wear, including blistering, graining, cuts, or debris.
    • Multimodal Fusion: We then intelligently fuse the outputs from both the telemetry predictions and the visual analysis. This creates a single, highly robust, real-time "Tire Health Index" that is more accurate than either data source alone.

  3. Decision Logic & Visualization \ This raw index is translated into actionable strategy. We are developing a decision engine to generate clear pit-stop recommendations, based on either exceeding a critical wear threshold or a more complex optimization model that balances track position against the cost of a stop. These insights will be delivered via an interactive React dashboard, displaying real-time wear estimates, optimal pit windows, and lap-by-lap performance trends.

  4. Validation & Refinement \ The entire system will be rigorously validated against historical race data. We will iteratively tune our models and fusion logic for maximum accuracy and refine the dashboard UI/UX for clarity and intuitive use under pressure.

Our Hackathon MVP Focus

We have a clear, strategic plan for the hackathon:

  • Build the core pipeline using pre-recorded historical telemetry and video.
  • Implement the foundational telemetry and visual detection models.
  • Develop the initial interactive dashboard to visualize the outputs.
  • As time permits, integrate the advanced multimodal fusion and optimization logic.

Challenges We Are Solving

We thrive on complex challenges. Here’s what we're tackling head-on:

  • Data Synchronization: Public datasets lack perfectly aligned telemetry and video. We are engineering a robust process to align these disparate data sources on a single timescale.
  • Real-Time Simulation: In the absence of a live F1 data feed, we are creatively developing a simulation environment to test our system's real-time performance.
  • Accuracy vs. Latency: We are balancing the power of deep learning with the need for real-time results, optimizing our models for high-speed inference without sacrificing predictive accuracy.
  • Inferential Labeling: Tire health is an inferred state, not a simple data point. We are applying advanced feature engineering to create reliable labels from multiple indicators (lap time drop-off, temperature spikes, visual wear).
  • Advanced Multimodal Fusion: Merging numerical time-series data with visual features is a complex challenge. We are implementing advanced fusion strategies to ensure the combined output is truly greater than the sum of its parts.
  • Model Generalizability: We are designing our models to be robust, aiming to avoid overfitting on a single track or condition, with the goal of generalizing across different circuits and weather scenarios.
  • Explainable AI (XAI): A "black box" is useless to a strategist. We are focused on building an interpretable dashboard that clearly communicates why a pit window is being recommended.
  • Resource Optimization: We are architecting our solution to run efficiently, ensuring continuous video analysis and ML inference are feasible even on limited hackathon hardware.
  • Actionable UX Design: The final challenge is translating complex AI insights into a clean, intuitive, and instantly actionable dashboard.

What Makes PitPerfect Different: Our Edge

  1. True Multimodal AI (Vision + Telemetry) \ While existing tools focus on sensor data, our solution is unique. We are fusing onboard video analysis (what the eye can see) with telemetry (what the sensors report). This adds a "human-like" perception layer, allowing our AI to see blistering and graining, not just infer it from lap times.

  2. Dynamic & Adaptive Strategy \ Traditional race models are pre-computed. PitPerfect is designed to be a dynamic system that continuously updates its predictions during the race, adapting in real-time to changing track temperatures, driver behavior, or unexpected weather.

  3. Explainable & Intuitive Dashboard \ We aren't just building a model; we're building a tool for decision-makers. Our dashboard prioritizes visual interpretability, showing why a pit stop is recommended, along with degradation trends and confidence scores.

  4. Cross-Domain Scalability \ This concept is bigger than racing. The core engine—fusing sensor and visual data for predictive maintenance—is designed to be generalizable, with clear applications in fleet management, robotics, and industrial machinery.

  5. Accessible & Open Prototype \ Unlike proprietary, closed-door F1 tools, our MVP is being built with public datasets and open-source models. This makes our concept replicable, verifiable, and a valuable contribution to the community.

Share this project:

Updates