v2.0.0

This release makes a major API update to switch from cold-start to hot-start when solving of MPC QP problems. It meanwhile adds two examples (with live plots) taken from real-robot use cases:

• The open-loop MPC part of the LIPM walking controller: python examples/lipm_walking_controller.py
• A closed-loop MPC balancer for Upkie, which also runs on the real robots: python examples/cart_pole.py

The problem class, now called MPCProblem, allows for initial/goal state and target trajectory updates. A new MPCQP class is introduced for QP vector updates (cost vectors depend on the initial state and targets, while on LTI problems QP matrices don't depend on these changing quantities, a source of optimization for faster solve times).

The new API also introduces two new submodules:

• Live plots: a debugging tool for plotting MPC trajectories while running (uses the TkAgg matplotlib backend)
• Systems: we start gathering common systems that can be re-used outside of examples. For instance, CartPole shipped with this release is used in a closed-loop MPC balancer on Upkie.

Added

• Base class for exceptions raised by this library
• Documentation checks using ruff
• Exception: ProblemDefinitionError
• MPCProblem: setter for the initial state
• MPCProblem: target state trajectory for stage state cost
• MPCQP class to update cost vectors during execution
• Plan: first_input getter
• Plan: is_empty property
• Started ltv_mpc.live_plots submodule
• Started ltv_mpc.systems submodule

Changed

• Always add state-input inequalities (even when unfeasible)
• Don't assume a Problem is fully defined in constructor
• Initial and goal states are now keyword arguments of MPCProblem
• Refactorbuild_qp into a new MPCQP class
• Rename Problem to MPCProblem
• Rename Solution to Plan
• Rename stacked_inputs to just inputs in plans
• Rename stacked_states to just states in plans
• Solver keyword argument to solve_mpc is now mandatory
• Use problem class from qpsolvers internally
• Warn rather than raise an exception when initial state is unfeasible