How cydal/lewm-car-nav works
This repository instruments an external 3D car-navigation simulator to collect, validate, and inspect trajectory datasets, then trains and evaluates a vector-observation LeWM model. The main workflow runs collection through the CLI, records episodes using configured driving policies, validates and reads the resulting dataset, and exports it for Stage 1 training. The graph also preserves the documented pilot, privileged-state capture, RGB/beacon options, and evaluation branches. Some subsystem links are documented rather than visible in the sampled implementation; those are dashed where appropriate.
Main entry points: collector.py Trajectory collector · __main__.py CLI commands · pilot.py Pilot analysis · dataset.py Dataset views
Components 20
Data collection
- Collection config
lewm/config.py - Driving profiles
lewm/policies.py - Trajectory collector
lewm/collector.py - Ground-truth capture
lewm/state.py - Goal beacon option
lewm/config.py
Dataset and validation
- Dataset views
lewm/dataset.py - Episode files
lewm/dataset.py - Dataset validator
lewm/validate.py
Stage 1 modeling
- HDF5 exporter
stage1_vector/export_hdf5.py - Training arrays
stage1_vector/export_hdf5.py - Windowed batches
stage1_vector/fast_loader.py - Vector JEPA
stage1_vector/model.py - Training loop
stage1_vector/train.py - Model evaluation
stage1_vector/baseline_eval.py
Pilot and interface
- CLI commands
lewm/__main__.py - Pilot analysis
lewm/pilot.py
Other components
- Experimenter
- Car simulator
- Official LeWM code
- Weights and Biases
Connections 24
- Experimenter → CLI commands: runs
README.md - CLI commands → Trajectory collector: dispatches
lewm/__main__.py - CLI commands → Pilot analysis: dispatches
lewm/__main__.py - CLI commands → Dataset validator: dispatches
lewm/__main__.py - Trajectory collector → Collection config: uses
lewm/collector.py - Trajectory collector → Driving profiles: uses
lewm/collector.py - Trajectory collector → Ground-truth capture: records
lewm/collector.py - Trajectory collector → Car simulator: steps
lewm/collector.py - Trajectory collector → Episode files: writes
lewm/collector.py - Dataset views → Episode files: reads
lewm/dataset.py - Dataset validator → Dataset views: checks
lewm/validate.py - Pilot analysis → Trajectory collector: collects
lewm/pilot.py - Pilot analysis → Dataset validator: validates
lewm/pilot.py - Pilot analysis → Dataset views: inspects
lewm/pilot.py - Collection config → Goal beacon option: configures
README.md - HDF5 exporter → Dataset views: reads
stage1_vector/export_hdf5.py - HDF5 exporter → Training arrays: writes
stage1_vector/export_hdf5.py - Windowed batches → Training arrays: loads
stage1_vector/fast_loader.py - Training loop → Windowed batches: batches
stage1_vector/train.py - Training loop → Vector JEPA: trains
stage1_vector/train.py - Vector JEPA → Official LeWM code: extends
stage1_vector/model.py - Training loop → Weights and Biases: logs optionally
stage1_vector/train.py - Model evaluation → Windowed batches: loads
stage1_vector/baseline_eval.py - Model evaluation → Vector JEPA: evaluates
stage1_vector/baseline_eval.py