How valeoai/v-max works
V-Max extends Waymax for autonomous-driving motion-planning research. The principal workflow is to load scenarios, construct a wrapped simulation environment, run a learned or rule-based planner, and train or evaluate it using driving metrics. Its major branches include RL/IL algorithms, configurable neural encoders, rule-based policies, ScenarioMax SDC-path support, and visualization. The graph reflects documented capabilities and indexed imports; several implementation details, particularly unsampled trainer and dataset internals, remain unverified.
Main entry points: utils.py Evaluation utilities · inference.py Policy inference · training.py Training pipeline · sim_factory.py Environment factory
Components 25
Training and evaluation
- Training entry
vmax/scripts/training/train.py - Evaluation entry
vmax/scripts/evaluate/evaluate.py - Training utilities
vmax/scripts/training/train_utils.py - Evaluation utilities
vmax/scripts/evaluate/utils.py
Planner agents
- Agent transitions
vmax/agents/datatypes.py - Behavior cloning
vmax/agents/learning/imitation/bc/bc_factory.py - SAC learner
vmax/agents/learning/reinforcement/sac/sac_factory.py - PPO learner
vmax/agents/learning/reinforcement/ppo/ppo_factory.py - BC-SAC learner
vmax/agents/learning/hybrid/bc_sac/bc_sac_factory.py - Network factory
vmax/agents/networks/network_factory.py - Wayformer encoder
vmax/agents/networks/encoders/wayformer.py - Policy inference
vmax/agents/pipeline/inference.py - Training pipeline
vmax/agents/pipeline/training.py - IDM policy
vmax/agents/rule_based/idm.py - PDM planner
vmax/agents/rule_based/pdm.py
Simulation and data
- Scenario datasets
vmax/simulator/datasets.py - Environment factory
vmax/simulator/sim_factory.py - Environment wrappers
vmax/simulator/wrappers/base.py - SDC path support
vmax/simulator/wrappers/sdc_path.py - Feature extraction
vmax/simulator/features/extractor/vec_extractor.py - Waymax overrides
vmax/simulator/overrides/datatypes/observation.py - Simulation operations
vmax/simulator/operations.py
Metrics and visualization
- Driving metrics
vmax/simulator/metrics/collector.py - Visualization
vmax/simulator/overrides/visualization/viz.py
Other components
- Researcher
Connections 33
- Researcher → Training entry: starts training
README.md - Researcher → Evaluation entry: runs evaluation
README.md - Training utilities → Scenario datasets: loads scenarios
vmax/scripts/training/train_utils.py - Evaluation entry → Evaluation utilities: uses utilities
vmax/scripts/evaluate/evaluate.py - Evaluation utilities → Scenario datasets: loads scenarios
vmax/scripts/evaluate/evaluate.py - Evaluation utilities → Driving metrics: collects metrics
vmax/scripts/evaluate/utils.py - Evaluation utilities → Visualization: renders results
vmax/scripts/evaluate/utils.py - Evaluation utilities → Behavior cloning: selects algorithm
vmax/scripts/evaluate/utils.py - Evaluation utilities → SAC learner: selects algorithm
vmax/scripts/evaluate/utils.py - Evaluation utilities → PPO learner: selects algorithm
vmax/scripts/evaluate/utils.py - Evaluation utilities → BC-SAC learner: selects algorithm
vmax/scripts/evaluate/utils.py - Environment factory → Driving metrics: registers metrics
vmax/simulator/sim_factory.py - Environment factory → Environment wrappers: configures environment
vmax/simulator/sim_factory.py - Environment factory → SDC path support: adds path wrapper
vmax/simulator/sim_factory.py - Training utilities → Scenario datasets: configures data
vmax/scripts/training/train_utils.py - Network factory → Wayformer encoder: builds encoders
vmax/agents/networks/network_factory.py - Policy inference → Agent transitions: creates transitions
vmax/agents/pipeline/inference.py - Policy inference → Environment wrappers: steps environment
vmax/agents/pipeline/inference.py - Policy inference → Simulation operations: selects SDC
vmax/agents/pipeline/inference.py - Policy inference → IDM policy: dispatches policy
vmax/agents/pipeline/inference.py - Policy inference → PDM planner: dispatches policy
vmax/agents/pipeline/inference.py - PDM planner → IDM policy: uses policy logic
vmax/agents/rule_based/pdm.py - Environment wrappers → Driving metrics: computes rewards
- Environment wrappers → Simulation operations: uses operations
vmax/simulator/wrappers/reward.py - Feature extraction → Simulation operations: uses operations
vmax/simulator/features/extractor/vec_extractor.py - Training entry → Environment factory: constructs simulation
- Training entry → Training pipeline: runs learning
- Training pipeline → Behavior cloning: trains policy
- Training pipeline → SAC learner: trains policy
- Training pipeline → PPO learner: trains policy
- Training pipeline → BC-SAC learner: trains policy
- SDC path support → Scenario datasets: uses enriched data
- Waymax overrides → Visualization: extends visualization
vmax/simulator/overrides/__init__.py