Skip to content

Examples

The examples/ directory contains robotics-related OCPs built with SSQP, covering offline trajectory generation as well as real-time MPC in simulation and on real systems.

See the Gallery for videos of what SSQP produces →

Directory Contents
trajgen/ Offline trajectory generation, in both Python and C++: SSQP generates a full trajectory from scratch given costs/constraints and little to no initialization. Includes di.py (the Quickstart double integrator), pendulums, the Panda and iiwa arms, and legged-robot examples (Go2, G1, A1, Talos). Most Python examples take command line parameters, and each one prints the final solver state and writes its result to a JSON file.
mujoco_mpc/ MPC examples using MuJoCo as the simulated "real" system: MuJoCo provides state feedback, and SSQP's MPC helper solves and re-solves the OCP each control step. Covers the pendubot (with optional online model identification), the Panda and the Go2. The simulation is synced with the controller, so it does not run in real time.
real_mpc/ MPC examples running against real hardware instead of a simulator: a double pendulum, a Franka Emika Panda and a Unitree Go2. Several of them use a server-client architecture, since the machine controlling the robot is not always the one running SSQP.
visualizer/ visualize.py / visualize_srbd.py, using viser to visualize trajectories written by solver.write(...).
models/ Custom SSQP StageDynamics/etc. models usable in place of XML files (the single rigid body model of a quadruped, and the planner that generates its contact schedules and references), plus related utilities.
assets/ URDF/MJCF assets used by the examples. The URDFs of SSQP's built in model presets (Go2, A1, Panda, G1, Talos and the double pendulum) come from here, and are embedded into the library when it is built.

Each directory has its own README with the details of every example, including the parameters it takes.

Solver benchmarks, comparing SSQP against other OCP solvers on a shared set of problems, live separately in benchmarks/ rather than under examples/; see Benchmarks.

Walkthroughs

A few examples are covered in detail:

  • Pendubot Swing-Up: trajgen, XML-based model, predefined costs, and a swing-up scheduling trick, ForwardDynamics mode

  • Panda Circle (Trajectory Generation): trajgen, frame-relative pose tracking, a real box constraint, InverseDynamics mode

  • Go2 Jump: trajgen, per-stage contact scheduling for a jump (stance → flight → stance), ReducedInverseDynamics mode

  • Panda Circle (MPC): mujoco_mpc, the same circle tracked in closed-loop MPC against a live MuJoCo simulation, with per-step target retargeting, ForwardDynamics mode by default (--inverse-mode switches to InverseDynamics)

Running the Python examples

The built ssqpy Python extension must be importable; see Using the Python bindings in Installation. Once built, source examples/setup_env.sh:

cd examples
source setup_env.sh
python trajgen/python/di.py

Then run any example script from its directory the same way (for example, cd trajgen/python/go2 and python go2_gaits.py --gait trot). The scripts take their settings as command line parameters, so --help lists them. Several example subdirectories (MuJoCo, real robot SDKs, the visualizer) use their own Python virtual environments for extra dependencies, so check for a *-venv/*venv directory or a README alongside the specific example before running it.

Running the C++ examples

The C++ trajectory generation examples are only built when the library is configured with examples enabled:

cmake -S . -B build -DBUILD_EXAMPLES=ON
cmake --build build

Each one becomes an executable under build/bin, in the same subfolder as its source, and should be run from the repository root, since asset and output paths are relative to it (for example, ./build/bin/urdf/panda/panda_circle).