mav_trajectory_generation_lib

ROS-free C++17 facade around ETH-ASL mav_trajectory_generation (degree-10 polynomial trajectories, Linear closed-form or NLopt nonlinear solver), with optional pybind11 bindings and a self-contained CLI example. Released under BSD-3-Clause.

Overview

mav_trajectory_generation_lib plans smooth polynomial trajectories (degree 10) through 3D waypoint sequences. The closed-form linear solver uses the segment-time allocation only; the NLopt-backed nonlinear solver additionally enforces velocity and acceleration magnitude bounds. The public API is built around a single one-shot entry point TrajectoryGenerator(GeneratorConfig) common to the trajectory-generation library family. Frame is ENU, units are SI, trajectory time always starts at t = 0. Sampled CSV output is the canonical 10-column schema t,px,py,pz,vx,vy,vz,ax,ay,az; the underlying spline is serialisable to <base>.spline.csv.

Repository layout

.
├── CMakeLists.txt
├── cmake/                                          # FetchContent shims for glog and NLopt
├── include/mav_trajectory_generation_cpp/          # Public headers (stable API)
│   ├── types.hpp                                   # POD configs (GeneratorConfig, OptimizationConfig, ...)
│   └── trajectory_generator.hpp
├── src/                                            # Implementation (Pimpl)
├── example/                                        # End-to-end CLI demo (utils + run_example.{cpp,py})
├── tests/                                          # GoogleTest, auto-discovered *_gtest.cpp
├── pybind/                                         # pybind11 module + Python package mirror (`mav_trajectory_generation_py`)
└── mav_trajectory_generation/                      # Vendored upstream submodule (immutable)

Dependencies

• C++17 compiler (GCC ≥ 9 / Clang ≥ 10).
• CMake ≥ 3.16.
• Eigen 3.
• yaml-cpp (only when BUILD_EXAMPLES=ON).
• glog and NLopt. If not installed system-wide, both are auto-fetched via CMake FetchContent (see cmake/).
• GoogleTest (C++ test suite).
• pybind11 ≥ 2.9 + Python ≥ 3.8 + numpy (Python bindings).
• PyYAML, matplotlib (used by example/run_example.py and example/plot_results.py).

Ubuntu / Debian:

sudo apt install build-essential cmake libeigen3-dev libyaml-cpp-dev \
                 libgoogle-glog-dev libnlopt-dev \
                 libgtest-dev pybind11-dev \
                 python3-numpy python3-yaml python3-matplotlib python3-pytest

Clone

Clone recursively to fetch the vendored upstream submodule:

git submodule update --init --recursive

Build

cmake -S . -B build \
      -DCMAKE_BUILD_TYPE=Release \
      -DBUILD_TESTING=ON \
      -DBUILD_EXAMPLES=ON \
      -DBUILD_PYBIND=ON
cmake --build build -j

CMake options:

Option	Default	Description
BUILD_TESTING	ON	Build and register the GoogleTest suite
BUILD_EXAMPLES	OFF	Build the C++ example binary
BUILD_PYBIND	OFF	Build the pybind11 module from the root
BUILD_DEVELOPER_TESTS	OFF	Enable lint/format/coverage checks
CMAKE_BUILD_TYPE	Release	Use Release for realistic solver timings

Install the Python package

Development (no pip install required): the build step with BUILD_PYBIND=ON stages the package under build/pybind/python/. Add it to your path:

export PYTHONPATH=$PWD/build/pybind/python:$PYTHONPATH

The example launchers export this path automatically.

Pip install (for use as a library in another project):

pip install ./pybind          # standard
pip install -e ./pybind       # editable

The wheel is built with scikit-build-core and is self-contained (RPATH=$ORIGIN).

Tests

C++

ctest --test-dir build --output-on-failure

Python

Ensure the package is in sys.path (see Install the Python package), then:

python3 -m pytest pybind/tests/ -q

Run the C++ example

./example/run_example_cpp.sh

Outputs example/trajectory_logs/trajectory.csv (10 cols t,px,py,pz,vx,vy,vz,ax,ay,az) plus the spline CSV example/trajectory_logs/trajectory.spline.csv and renders plot_3d.png

• plot_time.png.

Run the Python example

./example/run_example_py.sh

Same CSV / plots written under example/trajectory_logs/ with the _py suffix on the file names. The launcher exports build/pybind/python/ on PYTHONPATH automatically.

C++ usage

#include <vector>
#include <Eigen/Core>
#include <mav_trajectory_generation_cpp/trajectory_generator.hpp>

using namespace mav_trajectory_generation_cpp;

GeneratorConfig cfg;
cfg.optimization.solver = Solver::Nonlinear;   // NLopt; enforces v/a bounds
cfg.optimization.a_max  = 4.0;                  // [m/s^2]

TrajectoryGenerator generator(cfg);
std::vector<Waypoint> waypoints = {
    Waypoint(Eigen::Vector3d(0.0, 0.0, 1.0)),
    Waypoint(Eigen::Vector3d(5.0, 2.0, 1.5)),
    Waypoint(Eigen::Vector3d(8.0, 0.0, 2.0)),
};
if (!generator.generate(waypoints, /*max_speed=*/3.0)) {
  throw std::runtime_error("Trajectory generation failed");
}
const auto sample = generator.evaluate(0.5 * generator.duration());
// sample.position, sample.velocity, sample.acceleration, sample.jerk

Link against the mav_trajectory_generation_cpp CMake target (e.g. target_link_libraries(my_app PRIVATE mav_trajectory_generation_cpp)).

Python usage

import numpy as np
from mav_trajectory_generation_py import (
    GeneratorConfig, Solver, TrajectoryGenerator, Waypoint,
)

cfg = GeneratorConfig()
cfg.optimization.solver = Solver.nonlinear      # lower-case in Python
cfg.optimization.a_max  = 4.0                   # [m/s^2]

generator = TrajectoryGenerator(cfg)
waypoints = [
    Waypoint(np.array([0.0, 0.0, 1.0])),
    Waypoint(np.array([5.0, 2.0, 1.5])),
    Waypoint(np.array([8.0, 0.0, 2.0])),
]
assert generator.generate(waypoints, 3.0)
sample = generator.evaluate(0.5 * generator.duration())
# sample.position, sample.velocity, sample.acceleration, sample.jerk

Conventions

• Frame: ENU (X east, Y north, Z up).
• Units: SI (m, m/s, m/s², m/s³, s, rad, kg, N).
• Trajectory time always starts at t = 0.
• After failed generate(), every accessor throws std::runtime_error (C++) / RuntimeError (Python).
• CSV schema: 10 columns t,px,py,pz,vx,vy,vz,ax,ay,az. Spline serialisation in <base>.spline.csv with metadata header.

Repo-specific extras

• Solver selector Solver::Linear / Solver::Nonlinear (Solver.linear / Solver.nonlinear in Python). The linear path is closed-form least-squares; only the nonlinear path enforces velocity and acceleration magnitude bounds via NLopt.
• setSpline(const Spline&) — replay a previously serialised spline without re-running the solver. Useful for offline-planned trajectories fed to the same evaluator.
• Python wrapper mav_trajectory_generation_py.Trajectory (pybind/python/mav_trajectory_generation_py/trajectory.py): high-level helper that builds the native GeneratorConfig from a PyGeneratorConfig dataclass parsed from YAML (load_generator_config).

License

BSD 3-Clause — see LICENSE. Copyright © 2026 Universidad Politécnica de Madrid.

The vendored upstream algorithmic core under mav_trajectory_generation/ is authored by the ETH-ASL team (Achtelik, Burri, Oleynikova, Bähnemann, Popović) and licensed under Apache-2.0. BSD-3-Clause and Apache-2.0 are compatible: distributions of this repository must preserve both notices.

If the upstream submodule ever adds or renames .cpp files, the list MTG_CORE_SRC in CMakeLists.txt must be updated to keep the facade in sync.