MPC balancer

Make an Upkie to stand upright by closed-loop model predictive control.

Installation

We recommend using Anaconda to install the agent and all its dependencies in a clean environment:

conda env create -f environment.yaml
conda activate mpc_balancer

Alternatively, you should be able to install the packages listed in the environment file from PyPI.

Usage

To run in simulation, clone the upkie repository and run:

./start_simulation.sh

Activate your conda environment and run the agent by:

python run_agent.py

Solvers

This agent only works with QP solvers that support warm starting. At present we only support one solver:

Solver	Algorithm	License	Warm-start
ProxQP	Augmented Lagrangian	BSD-2-Clause	✔️

You can take a peek at the ProxQP balancer (research code) for more solvers.

Export dependencies to your Upkie

This agent can export a pixi environment to your Upkie using pixi-pack. If you don't have pixi yet, you will need to install it first. Then, to pack an environment from your computer, run:

make pack_env

This will create an environment.tar archive in the current directory. You can upload it to your Upkie by make upload and unpack it from the agent's remote directory by:

your_user@your_upkie:~/mpc_balancer$ make unpack_env

If pixi-pack is not installed on your Upkie, you can download the pixi-pack-aarch64-unknown-linux-musl binary from the pixi-pack release page. Finally, activate the environment and run the agent:

conda activate mpc_balancer
python mpc_balancer/run.py

See also

• PPO balancer: an MLP agent trained for the same task by reinforcement learning.
• ProxQP balancer: prototype for this agent used in the code for the ProxQP paper. Currently supports more QP solvers.