Katkat
Back

How predict vehicles’ turn direction at intersection?

WaymoData Scientist · Onsite
EasyData ScientistOnsiteMachine Learning

Last updated

This problem tests the ability to design an end-to-end machine learning pipeline for vehicle intent prediction, focusing on feature engineering, sequence modeling, and temporal validation.

Scenario

You are working on a self-driving perception stack at Waymo. Your task is to build a model that predicts whether an approaching vehicle at a 4-way intersection will turn left, turn right, or continue straight. You have access to a dataset containing high-frequency GPS, velocity, and acceleration logs for vehicles within a 50-meter radius of the intersection, as well as static HD map data (lane centerlines, stop lines, and traffic light states).

Task

  1. Label Definition: How would you define the 'ground truth' target for a vehicle that is currently in the middle of an intersection? How do you handle ambiguous cases where a driver slows down but hasn't committed to a turn?
  2. Feature Engineering: Propose a set of features that combine the vehicle's recent trajectory with map constraints. How do you represent the relationship between the vehicle's position and the intersection's lane geometry?
  3. Model Selection: Given the sequential nature of the data, which model architecture would you choose (e.g., RNNs, Transformers, or Graph Neural Networks) and why?
  4. Evaluation: Explain how you would structure your training/validation/test split to prevent temporal leakage, given that vehicle maneuvers are highly correlated within the same intersection over time.

Constraints

  • The system must provide a prediction with at least 500ms of lead time before the vehicle reaches the intersection center.
  • The model must be robust to noisy GPS sensor data.

Hints

Hint 1

Consider defining the label based on the vehicle's exit lane rather than instantaneous steering angle.

Hint 2

Think about using relative distance and heading angle to the nearest lane centerline as a spatial feature.

Hint 3

Ensure your data splitting strategy groups by intersection ID or recording session to avoid overfitting to specific geographic layouts.

Follow-up questions

  • How would you handle the 'multi-modal' nature of the output if a vehicle could potentially go straight or turn right from the same lane?
  • If the model performance drops significantly at night or during rain, what data augmentation techniques would you suggest?
  • How would you integrate the current traffic light state into your model, and how would that influence the predicted intent?