A modular and extensible Python port of the TUD-AMR MPC Planner, designed for real-time motion planning under constraints. PyMPC brings together math utilities, solver-agnostic optimization backends (e.g. CasADi, OSQP), and a flexible constraint interface to support rapid prototyping and deployment of Model Predictive Control systems.
- Model Predictive Control (MPC) framework in Python
- Modular design for swapping solvers or models
- Support for:
- State/input constraints
- Soft and hard constraints
- Obstacle avoidance
- Built-in math utilities for dynamics, linearization, etc.
- Modification of the original C++ codebase
- Unit-test friendly structure for rapid development
| Constraint Type | Main Idea | Obstacle Handling | Uncertainty Handling | Computational Cost | Compatible Solvers |
|---|---|---|---|---|---|
| LinearizedConstraints | Linearize nonlinear constraints near current trajectory | Deterministic positions | ❌ No | Low | QP, SQP |
| EllipsoidConstraints | Approximate obstacles as ellipsoids for smooth convex regions | Static or dynamic obstacles | ❌ No | Medium | SQP, NLP |
| GaussianConstraints | Probabilistic modeling using Gaussian mean & covariance | Dynamic obstacles with Gaussian uncertainty | ✅ Yes | Medium-High | NLP |
| DecompositionConstraints | Divide road into convex safe regions | Road boundaries, static obstacles | ❌ No | Medium | QP, SQP |
| GuidanceConstraints | Multi-homotopy guided optimization (parallel planners) | All obstacle types via global guidance | ✅ Partial (via global plan diversity) | High | NLP (or multiple QPs in parallel) |
| ScenarioConstraints | Robust MPC via sampled uncertainty scenarios | Requires probabilistic predictions (non-deterministic) | ✅ Yes | Very High | NLP |
| Method | How It Handles Uncertainty | Key Parameters | Pros | Cons |
|---|---|---|---|---|
| GaussianConstraints | Chance constraints using Gaussian distributions | Risk level (α), covariance of obstacle | Smooth, risk-adjustable, mathematically elegant | Requires accurate uncertainty model; heavier solve |
| ScenarioConstraints | Monte Carlo sampling of obstacle trajectories | Number of scenarios, safe horizon | Robust to arbitrary distributions | Very expensive, needs many solvers in parallel |
| GuidanceConstraints | Uses multiple global guidance paths to hedge risk | Number of homotopy classes (paths) | Finds safe topologies, avoids local minima | Indirect uncertainty handling, depends on guidance planner |
git clone https://github.com/stephen-crawford/PyMPC.git
cd PyMPC
pip install -e .
Requires Python 3.8+, NumPy, and optionally CasADi or OSQP depending on your backend.To run tests you may need to add the following run configuration: LD_LIBRARY_PATH=/opt/ros/humble/lib:/opt/ros/humble/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH
Related Work This library is a Python port of the excellent tud-amr/mpc_planner, originally written in C++. Our goal is to preserve its structure and intent while providing a more flexible, Pythonic interface for rapid development and experimentation.
Contributing Contributions welcome! If you'd like to add models, constraints, or solver support, open a pull request or issue.
MIT License