Overview
What this challenge is about.
Use the Argoverse 2 motion-forecasting dataset (open access). Train an LSTM baseline + a transformer challenger (e.g., a small Wayformer or HiVT). Evaluate on minADE/minFDE (minimum average and final displacement error over k=6 trajectories) overall and on the unprotected-left-turn interaction slice. Report training cost and inference latency on a single A10 GPU. Deliver a 4-page research memo with the go/hold call.
The Brief
What you'll do, and what you'll demonstrate.
Decide whether to invest a quarter of behavior-team engineering into the transformer predictor over the LSTM baseline, based on accuracy + cost evidence.
Earning criteria — what you'll demonstrate
- Implement and tune trajectory predictors on a real AV dataset
- Define and evaluate behavior-relevant interaction slices
- Trade off accuracy gains against training and inference cost
- Write a research memo a behavior team can decide on
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
ML Researcher
Rigorous, slice-aware benchmarking of trajectory predictors is the daily work of a behavior-team ML researcher at any AV company.
This challenge sharpens
- trajectory-prediction
- transformer-models
- experiment-design
Research Scientist
Honest reporting of cost + accuracy trade-offs across architectures is the research-scientist discipline AV labs hire for.
This challenge sharpens
- transformer-models
- evaluation
- benchmarking
Machine Learning Engineer
Pipeline + reproducibility habits learned here transfer directly into the MLE side of behavior-prediction productionization.
This challenge sharpens
- pytorch
- benchmarking
- trajectory-prediction