Skip to main content
Back

UT Austin's FE-DPC Beats Industrial Controllers While Running 71x Faster

A new control framework blends function encoders with differentiable predictive control, letting policies adapt to unseen nonlinear dynamics without any retraining or online optimization.

  • FE-DPC combines function encoders and differentiable predictive control for zero-shot adaptive control of nonlinear systems: arxiv.org/abs/2511.05757
  • Represents unknown dynamics as linear combinations of learned neural ODE basis functions with coefficients estimated online.
  • Policy conditioned on coefficients adapts instantly without retraining or online optimization.
  • Achieves 2.3x to 71.5x speedup over classical MPC across four benchmarks with comparable tracking accuracy.
  • Tested on Van der Pol, two-tank, 7D glycolytic oscillator, and 12D quadrotor with mid-simulation dynamics switches.
  • Open-source code available at github.com/hassaniqbal209/DPCFunctionEncoder

Model predictive control has been the workhorse of industrial control for decades, but it carries a heavy price. Every time the plant it controls changes, the controller has to re-solve an optimization problem online, and every time the underlying dynamics drift, the model has to be re-identified. A new paper from researchers at UT Austin and Johns Hopkins proposes a way to sidestep both bottlenecks by pre-training a single neural policy that adapts instantly to a whole family of nonlinear systems.

The method, called Function Encoder-based Differentiable Predictive Control (FE-DPC), combines two recent ideas: function encoders that represent unknown dynamics as a linear combination of learned neural ODE basis functions, and differentiable predictive control that trains explicit policies by backpropagating through a differentiable simulator. Across four nonlinear benchmarks, the controller matches white-box MPC on tracking accuracy while running up to 71 times faster.

Why MPC keeps getting stuck

MPC remains the industry standard for controlling multi-input, multi-output constrained systems, and it excels at accuracy and reliability through receding-horizon optimization. As control problems grow more complex, though, repeatedly solving those online optimizations with accurate models can become prohibitively expensive. That is the first pain point. The second is that classical MPC assumes an accurate, differentiable model of the plant, which rarely survives contact with reality once sensors drift, payloads shift, or wind fields change.

Learning-based alternatives have tried to fix this by amortizing the computation offline, but most of them bake in a specific set of dynamics. Change the mass of a drone or the valve coefficient of a tank, and you are back to retraining. Function encoders offer retraining-free adaptability across dynamics that existing methods often lack.