leap-c

leap-c or (learning predictive control) provides a simple way how to wrap optimal-control solvers into modern deep learning pipelines like PyTorch.

This repository provides a differentiable layer based on the very fast acados framework. Higher-level planners, environments, and RL training utilities live in downstream projects such as leap-c-lab and mpc-sac.

Note

The repository recently moved toward a smaller core interface. The previous broader interface is available in v0.2.0-alpha.

Key Features

• A simple torch interface for acados called AcadosDiffMpcTorch.
• Solve optimal control problems in parallel using multithreading.
• Backpropagate exact solution sensitivities through the MPC layer.
• Retrieve sensitivities of the optimal cost and the optimal control sequence with respect to problem parameters.
• Details as warm-starting, initialization and more are conveniently handled by the AcadosDiffMpcTorch interface.

Installation

Follow the installation guide.

Documentation

• Getting started
• Define a differentiable MPC
• Parameter management
• Running notebooks
• API reference

Minimal Interaction

import torch

from leap_c.torch import AcadosDiffMpcTorch

# Build these with acados/CasADi and leap-c's parameter manager.
ocp = build_acados_ocp(...)
parameter_manager = build_parameter_manager(...)

diff_mpc = AcadosDiffMpcTorch(ocp, parameter_manager)

x0 = torch.tensor([[1.0, 0.0]], dtype=torch.float64)
Q = torch.tensor([[1.0, 1.0]], dtype=torch.float64, requires_grad=True)

ctx, u0, x, u, value = diff_mpc(x0, params={"Q": Q})
value.sum().backward()

print(u0)      # first optimal action
print(Q.grad)  # gradient through the MPC solve

The snippet omits the acados OCP construction. For a full runnable example, see Define a differentiable MPC.

Ecosystem

• leap-c-lab: example environments, OCP definitions, and planners that show how to build on the core layer.
• mpc-sac: SAC/CrossQ-style RL training utilities for learning MPC controllers with leap-c and leap-c-lab.

Questions?

Open a new thread or browse existing ones on the GitHub discussions page.

Citing

If you use code from this repository in your work, please cite:

@misc{fichtner_leapc_2025,
  title = {Leap-c/Leap-c: V0.1.0-Alpha},
  author = {Fichtner, Leonard and Reinhardt, Dirk and Hoffmann, Jasper and Airaldi, Filippo and Frey, Jonathan and Kir Hromatko, Josip and Baumgaertner, Katrin and Amria, Mazen and Reiter, Rudolf and Sawant, Shambhuraj},
  year = 2025,
  month = oct,
  howpublished = {Zenodo}
}

Related Projects

The following projects follow similar ideas and might be interesting:

• mpc.pytorch: Early work on embedding MPC in PyTorch for end-to-end learning, with a more restricted class of MPC problems
• mpcrl: A simpler codebase for using RL with MPC as function approximator
• Neuromancer: A differentiable programming library that allows to include parametric optimization layers (including MPC) in PyTorch computational graphs
• ntnu-itk-autonomous-ship-lab/rlmpc: A codebase tailored for marine vessel control using RL and MPC