How tumftm/ethicaltrajectoryplanning works
Project type and purpose
• This is a Python-based autonomous vehicle trajectory planning tool with an ethical risk-assessment core. It combines two planning layers (GlobalPath + Frenet), a risk assessment module, configuration management, prediction networks, and a suite of analysis/visualization scripts. It runs as a CLI application on CommonRoad scenarios, producing logs, plots, and trajectories.
Main components
a. Inputs & Configuration
• CommonRoad scenario repository (external folder)
• JSON config files under planner/Frenet/configs
b. Core Planning Layers
• GlobalPath planner (planner/GlobalPath)
– Lateral/longitudinal path generation via quintic polynomials
• Frenet corridor planner (planner/Frenet)
– Path sampling + trajectory cost evaluation
– Prediction network (GPU-enabled) for dynamic obstacles
– Cost terms including collision risk and harm estimation
– Bayesian optimization for parameter tuning (plannertools)
c. Risk Assessment Module (risk_assessment)
• collision_probability, harm_estimation, risk_costs
• helper functions & logistic regression models
• visualization dashboards (risk_assessment/visualization)
d. Tools & Utilities
• Analysis tools (planner/Frenet/analyze_tools)
• Plotting tools (planner/Frenet/plot_tools)
• Scenario handler, animate, evaluate, write_solution (planner/plannertools)
• Common utilities (planner/utils, risk_assessment/utils)
e. Output & Storage
• Logs (planner/Frenet/results)
• Evaluation metrics (JSON, boxplots, dashboards)
Relationships & Data Flow
1. Load CommonRoad scenario →
2. Load configs →
3. GlobalPath planner produces rough path →
4. Frenet planner samples trajectories around rough path →
• Calls prediction network
• Invokes risk_assessment (collision & harm) to compute cost
• Optionally runs Bayesian optimization to tune weights
5. Select best trajectory → write_solution →
6. Log results →
7. Post–processing: analyze_tools → plot_tools → final metrics/plots
Architectural patterns & design principles
• Modular subpackages for separation of concerns (planner.Frenet, planner.GlobalPath, risk_assessment)
• Configuration‐driven behavior (JSON files)
• Script‐based orchestration (entry scripts like frenet_planner.py)
• Reusable utilities & helper libraries
• GPU‐accelerated prediction as an external “service”
Essential elements
• Component labels (e.g., “planner/GlobalPath”, “risk_assessment/collision_probability”)
• On console/CLI entry points (frenet_planner.py, evaluatefrenet.py)