🅿️ MPC Parking Assistant
MPC Parking Assistant
Interactive 3D visualization of Model Predictive Control (MPC) applied to autonomous vehicle parking.
Overview
This project demonstrates how robots "look ahead" to plan optimal trajectories while respecting physical constraints. Unlike reactive PID controllers, MPC predicts future states and optimizes control inputs over a time horizon.
Key Features
Optimization-Based Control:
- Gradient descent solver with shooting method
- Real-time trajectory generation (re-planned every step)
- Tunable cost function weights (position, angle, velocity, control effort)
Non-Holonomic Vehicle Physics:
- Bicycle model kinematics
- Ackermann steering constraints
- Cannot move sideways - must perform complex maneuvers to park
Interactive Modes:
- Manual Mode: Drive with WASD/arrow keys to feel the physics
- Autopilot Mode: AI solver plans and executes parking trajectory
Visualization:
- Ghost cars showing prediction horizon
- Actual vs. predicted path
- Real-time performance metrics (solver time, distance to target, cost)
How It Works
The MPC Algorithm
- Predict: Roll out vehicle trajectory over N future timesteps using bicycle model
- Optimize: Minimize cost function using gradient descent:
- Cost = position_error² + angle_error² + velocity² + control_effort²
- Execute: Apply first control input, shift horizon, repeat
Non-Holonomic Constraint
The bicycle model enforces: ẏcos(θ) - ẋsin(θ) = 0
This constraint prevents sideways motion, forcing the S-curve maneuver in parallel parking.
Cost Function Tuning
- Position weight: How much to prioritize reaching target location
- Angle weight: How much to prioritize correct orientation
- Velocity weight: Prefer slower parking (safer)
- Control weights: Penalize aggressive acceleration/steering (smoother)
Technical Implementation
Algorithm: Gradient descent shooting method with backtracking line search Physics: Euler integration of bicycle model (5 state variables, 2 control inputs) Solver Performance: ~10-50ms per solve on desktop, 20-100ms on mobile Device Adaptive: Reduces horizon length and ghost count on low-power devices
Try It
Use the controls to:
- Switch between Manual and Autopilot modes
- Adjust MPC horizon length (how far ahead the robot looks)
- Tune cost weights to see how they affect parking behavior
- Watch the ghost cars show the predicted future trajectory
Learning Points
- MPC vs PID: Predictive optimal control vs reactive error correction
- Non-holonomic systems: Why cars need complex maneuvers
- Optimization in real-time: Balancing solution quality with computation time
- Multi-objective optimization: Trading off competing goals (speed, smoothness, accuracy)
