Interpretable Autonomy: Using Mixture-of Experts for End-to-End Self-Driving

Usage:

1. Recording Data

Run src/carla_record.py viz to record a dataset with live visualizations, omit viz to run headless. You will need an instance of CARLA simulator running on the same computer.

2. Model Definitions

Models are defined in src/models.py. Relevant ones are WaypointNet and MultiCamWaypointNet (MoE coming soon).

3. Training Script

See src/train.py for the training script and arguments. Currently batchsize is set to 1 which may limit training speed.

4. Evaluation

Run carla_model_test.py viz to see evaluation of trained models in CARLA (closed-loop eval). Omitting viz will print a dictionary of failure events (crashes, invalid generated waypoints) to stdout after 100,000 steps of the simulation. For open-loop evaluation, see the entrypoint to src/utils.py waypoint.