Waymo·Machine Learning Engineer·Technical Phone Screen·Senior

Senior

May 2026

Summary

Waymo ML engineer interview with a problem I didn't fully expect: given raw 3D trajectory data, figure out how to preprocess and tokenize it so a model can actually predict the next step in the sequence. More engineering judgment than pure ML theory.

Questions Asked (1)

You're given 3D trajectory data. How would you preprocess and tokenize it into a format suitable for predicting the next step in the sequence?

Data ModelingSystem DesignTechnical Trade-offs

Author's notes

This one took me a second to even figure out where to start.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by outlining the key preprocessing steps: cleaning, coordinate transformation, and normalization. Then describe tokenization strategies such as discretizing space into a grid or using a learned vocabulary, and discuss how to structure the sequence for next-step prediction. Emphasize trade-offs between resolution, sequence length, and model complexity.

Pro tip: Mention that in autonomous driving, preserving temporal consistency and handling missing data are critical; propose using relative coordinates and velocity features to improve generalization across scenes.

1. Data Cleaning and Filtering

Remove outliers, handle missing values via interpolation, and filter noise using techniques like Kalman filters or moving averages.

2. Coordinate Transformation and Normalization

Convert absolute coordinates to a local frame (e.g., ego-vehicle centric) and normalize features (position, velocity, acceleration) to zero mean and unit variance.

3. Tokenization Strategy

Discretize continuous space into tokens using grid cells, clustering, or learned embeddings; consider using a vocabulary of motion primitives or relative displacements.

4. Sequence Construction

Create input-output pairs by sliding a window over the tokenized sequence, ensuring fixed-length context and next-step target; handle variable-length sequences with padding or masking.

5. Model-Specific Considerations

Choose tokenization granularity based on model capacity and prediction horizon; evaluate trade-offs between coarse tokens (longer sequences) and fine tokens (shorter sequences).

Key Points to Mention

Handling missing data and outliers in trajectory data
Coordinate frame normalization (e.g., ego-centric) for translation and rotation invariance
Discretization methods: grid-based, clustering, or learned vector quantization
Feature engineering: including velocity, acceleration, and heading
Sequence windowing and padding for batch training
Trade-offs between token resolution, sequence length, and model performance

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.