Differentiable Model Predictive Control

[avik-pal]

Box Constrained MPC can be expressed as an LCP (See Linear complementarity model predictive control with limited iterations for box-constrained problems for the exact mathematical details.

LCP/MCP algorithms are simpler to batch and "gpuify" than most QP solvers out there.

• Implement MPC using LCPs
• Differentiable MPC
• Compare against https://github.com/locuslab/qpth and https://github.com/locuslab/mpc.pytorch

[avik-pal]

Related packages:

• Direct Methods: https://github.com/thowell/DirectTrajectoryOptimization.jl
• QPTH julia version using Jump: https://jump.dev/DiffOpt.jl/stable/manual/
• Iterative LQR: https://github.com/thowell/IterativeLQR.jl
  • https://github.com/thowell/optimization_dynamics