⚡ electrical engineering

Zero-Shot Function Encoder-Based Differentiable Predictive Control

This paper proposes a differentiable framework that combines function encoder-based neural ODEs with differentiable predictive control to enable efficient, zero-shot adaptive control of nonlinear dynamical systems without the need for retraining or costly online optimization.

Original authors: Hassan Iqbal, Xingjian Li, Tyler Ingebrand, Adam Thorpe, Krishna Kumar, Ufuk Topcu, Ján Drgoňa

Published 2026-04-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hassan Iqbal, Xingjian Li, Tyler Ingebrand, Adam Thorpe, Krishna Kumar, Ufuk Topcu, Ján Drgoňa

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a pilot flying a drone. Usually, if the wind suddenly changes, or if you pick up a heavy package that shifts your balance, your drone's computer has to stop, recalculate everything, and figure out a new flight path. This takes time, uses a lot of battery, and if the change happens too fast, the drone might crash.

This paper introduces a new way to fly drones (and control other machines) that doesn't need to stop and recalculate. It's like giving the drone a super-intuitive sense of balance that works instantly, no matter what changes.

Here is the breakdown of how they did it, using simple analogies:

1. The Problem: The "Over-Thinker" vs. The "Instinct"

The Old Way (Model Predictive Control): Imagine a pilot who is a math genius. Before every move, they pull out a massive calculator, measure the wind, weigh the cargo, and solve a complex equation to decide which way to tilt. It's very accurate, but it's slow. If the wind changes while they are calculating, the answer is already wrong.
The New Way (Zero-Shot Adaptive Control): Imagine a pilot who has flown thousands of different drones in every possible weather condition. They don't need to calculate; they just feel the change and react instantly. This paper teaches a computer to have that same "instinct."

2. The Secret Sauce: The "Universal Translator" (Function Encoder)

The core of this invention is something called a Function Encoder. Think of this as a Universal Translator for machine behavior.

The Library of Movements: The researchers first taught the computer a "library" of basic movement patterns (like how a car turns, how a pendulum swings, or how a drone hovers). These are the "basis functions."
The Recipe Card: When a new machine appears (say, a drone with a broken motor or a different weight), the system doesn't need to learn how to fly it from scratch. Instead, it looks at the new machine for just a few seconds and says, "Ah, this new machine is just a mix of 10% of Pattern A, 50% of Pattern B, and 20% of Pattern C."
The Magic: It creates a tiny "recipe card" (called coefficients) that describes the new machine. It doesn't need to retrain the whole pilot; it just updates the recipe card.

3. The Pilot: The "Differentiable Predictive Control" (DPC)

Once the system has the "recipe card" (the coefficients), it hands it to the DPC Pilot.

The Pre-Trained Pilot: This pilot was trained offline (in a simulator) to handle every possible combination of those recipe cards.
Zero-Shot Adaptation: "Zero-shot" means the pilot has never seen this specific machine before, but because they understand the language of the recipe cards, they know exactly how to fly it immediately.
No Re-calculating: Unlike the math genius pilot, this pilot doesn't stop to solve equations. They just look at the recipe card, see the current state, and push the controls. It's instant.

4. The Results: Fast, Accurate, and Tough

The authors tested this on four very different "machines":

A Bouncing Spring (Van der Pol): A system that naturally wants to oscillate wildly. The new method stabilized it instantly, even when the spring's stiffness changed mid-flight.
Two Water Tanks: A system where water flows between tanks. Even when the valves changed size, the system kept the water levels perfect.
A Yeast Cell (Glycolytic Oscillator): A very complex, "stiff" biological system. The method handled it much faster than traditional computers could.
A Quadrotor Drone: The ultimate test. They tested 20 different drones with different weights and shapes. Even when the drone's weight suddenly changed mid-air, the system kept it hovering perfectly.

The Big Win:
In the tests, this new method was 2 to 70 times faster than the traditional "math genius" method (MPC), while still being almost as accurate. It proved that you don't need to be a slow calculator to be a good pilot; you just need a good intuition built on a solid foundation.

Summary

Think of this paper as teaching a robot to ride a bike.

Old way: The robot stops, measures the wind, checks the tire pressure, and calculates the physics of leaning before every turn.
New way: The robot learns the "feeling" of riding different bikes. When it gets on a new bike with a flat tire or a heavy backpack, it doesn't stop to calculate. It just adjusts its balance instantly based on the "feeling" (the coefficients) and keeps riding smoothly.

This makes robots and self-driving cars safer and faster because they can adapt to the real world instantly, without getting stuck in a loop of calculations.

1. Problem Statement

The paper addresses the challenge of adaptive control for nonlinear dynamical systems where system parameters change online or are initially unknown.

Limitations of Existing Methods:
- Model Predictive Control (MPC): While accurate, standard MPC requires solving expensive online optimization problems at every time step. It also relies on having an exact, fully differentiable model of the system dynamics, which is often unavailable in real-world scenarios due to noise, degradation, or uncertainty.
- Learning-Based Control: Many data-driven methods rely on fixed models or pre-trained datasets. They often fail to generalize when system parameters shift (e.g., changing payloads on a drone or wind conditions) without costly retraining or re-optimization.
Goal: The authors aim to develop a control framework that achieves zero-shot adaptation. This means the controller must adapt to new, unseen system dynamics in real-time using limited observations, without requiring online model retraining or expensive iterative optimization.

2. Methodology

The proposed framework, Function-Encoder Differentiable Predictive Control (FE-DPC), integrates two core components: Function Encoders (FE) based on Neural Ordinary Differential Equations (NODEs) and Differentiable Predictive Control (DPC).

A. System Identification via Function Encoders (FE-NODE)

Instead of learning a single fixed dynamics model, the authors model the family of possible system dynamics as a linear combination of learned basis functions.

Offline Training: A set of neural ODE basis functions $\{g_1, \dots, g_B\}$ is trained to span the space of possible system dynamics. These basis functions are parameterized by Neural Networks (MLPs).
Online Adaptation: For a new, unseen system, the specific dynamics are represented by a compact vector of coefficients $c$ . These coefficients are computed online in closed-form using regularized least squares (Tikhonov regularization) on a small amount of observed input-output data:
$(G + \lambda I)c = F$
where $G$ and $F$ are derived from inner products of the basis functions and observed data.
Key Advantage: The basis functions are frozen after offline training. Adapting to a new system only requires inferring the coefficients $c$ , which is computationally cheap and does not require gradient-based retraining of the neural network.

B. Differentiable Predictive Control (DPC)

DPC is used to learn an explicit control policy $\pi_W(x; \xi, c)$ that maps states and parameters to control actions.

Policy Conditioning: The control policy is conditioned on the system parameters $\xi$ (constraints/objectives) and the FE coefficients $c$ (dynamics).
Offline Training: The policy network is trained via backpropagation through the entire trajectory rollout (unrolling the FE-NODE dynamics). The loss function minimizes tracking error and control effort while penalizing constraint violations.
Zero-Shot Inference: Once trained, the policy can handle new dynamics simply by updating the input coefficient vector $c$ . No online optimization (like solving an MPC quadratic program) is required during execution.

C. Workflow

Offline: Train FE basis functions and the parametric DPC policy network.
Online (Adaptation):
- Collect a short trajectory of the new system.
- Compute the FE coefficients $c$ via least squares.
- Feed $c$ into the pre-trained policy network to generate control actions instantly.

3. Key Contributions

Zero-Shot Control Architecture: The authors propose a novel architecture where system dynamics are represented by a compact FE-NODE basis. This enables the controller to adapt to unseen dynamics instantly by inferring coefficients, eliminating the need for online retraining.
Scalable Performance on Complex Benchmarks: The method is validated on four diverse, nonlinear, and high-dimensional control tasks:
- Van der Pol Oscillator: Stabilization with switching dynamics.
- Two-Tank System: Level regulation with varying valve coefficients.
- Glycolytic Oscillator (GO): Control of a stiff, 7-dimensional biochemical system.
- Quadrotor: 12-dimensional flight control with varying mass and inertia.
Computational Efficiency: The approach achieves orders-of-magnitude faster inference compared to online MPC solvers while maintaining competitive control accuracy.
Open Source: The authors released the code, models, and training pipelines to ensure reproducibility.

4. Experimental Results

The paper evaluates FE-DPC against a "White-Box" (WB) MPC baseline (which assumes perfect knowledge of the true dynamics).

Accuracy:
- In the Van der Pol and Two-Tank scenarios, FE-DPC successfully stabilized systems and tracked references even when dynamics switched abruptly during simulation.
- In the Glycolytic Oscillator (a stiff system), FE-DPC achieved tracking performance comparable to WB-MPC, though with slightly higher error (MSE 0.18 vs 0.03) due to the complexity of the dynamics.
- In the Quadrotor task, the controller maintained stable hover across 20 distinct parameterizations and random dynamic switches.
Efficiency (Inference Time):
- FE-DPC demonstrated significant speedups over WB-MPC:
  - Van der Pol: 2.3x faster.
  - Two-Tank: 6x faster.
  - Glycolytic Oscillator: 23x faster (5.89s vs 136.07s).
  - Quadrotor: 80x faster (1.93s vs 155.85s).
Robustness: The method maintained closed-loop stability under abrupt parameter changes without any re-optimization, demonstrating true zero-shot adaptability.

5. Significance and Conclusion

This work bridges the gap between the high accuracy of model-based control (MPC) and the computational efficiency of learning-based control.

Practical Impact: It offers a solution for safety-critical and real-time applications (e.g., aerospace, robotics) where system parameters are uncertain or time-varying, and where the computational budget precludes online optimization.
Theoretical Contribution: It formalizes the use of Function Encoders to create a "universal" dynamics representation that can be linearly combined to approximate any system within a learned Hilbert space, providing finite-sample guarantees for approximation error.
Future Work: The authors note that while MLPs were used for the basis functions, other architectures (e.g., transformers or graph networks) could further improve performance, and the method relies on sufficient data coverage of the dynamics space during the offline phase.

In summary, FE-DPC provides a general-purpose, fast, and adaptive control tool that learns to control a family of systems offline and adapts to specific instances online with minimal computational overhead.