Learning Dynamics from Observations for Black Box Control

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 111 papers15 conceptsfrozen 2026-10-04

Traditional control theory relies on a "physics-first" approach: to control a vehicle or a power grid, you first write down the differential equations governing its motion, torque, and friction. However, as systems grow more complex—or when their internal logic is proprietary—this white-box modeling becomes a bottleneck.

A growing body of research suggests a different path: treating the system as a black box and learning its dynamics directly from input-output data. By observing how a system responds to specific commands, we can build "surrogate models" that are often more accurate and easier to deploy than their physics-based counterparts.

The Shift to Data-Driven Calibration

In industrial settings, manual tuning is the standard but expensive way to handle mechanical variations across a fleet. Fan Zhu and colleagues at Baidu demonstrated that this can be automated using the Baidu Apollo Auto-Calibration System.

Instead of modeling the intricate physics of engine torque or transmission loss, they used a feedforward neural network to map control inputs directly to observed acceleration. By training on just 20 minutes of human driving data, the system learns the specific "personality" of a vehicle, allowing for precise longitudinal control that adapts to mechanical wear and changing loads without human intervention.

Modeling the Unobservable

This data-driven philosophy is particularly powerful when the system's internal logic is literally hidden from the user. When Inkyu Sa and Mina Kamel sought to build a high-performance research drone using a closed-source DJI flight controller, they couldn't access the source code. Instead, they recorded the drone's physical responses to specific thrust and attitude commands. This empirical mapping allowed them to wrap a sophisticated nonlinear controller around a "black box" hardware unit, achieving centimeter-level precision.

A similar logic applies to auditing commercial technology. George Gunter and Raphael Stern used system identification to analyze commercial Adaptive Cruise Control (ACC). By mapping observed driving data to mathematical models, they could identify internal gains and time headways of proprietary systems, eventually proving that these systems are "string unstable" and contribute to traffic jams.

Beyond Regression: Neural ODEs and Transformers

Modern research has moved beyond simple curve-fitting to more sophisticated architectures that respect the temporal nature of physics.

  1. Neural ODEs: Vrushabh Zinage and Efstathios Bakolas introduced Neural Differentiable Integral Control Barrier Functions. This framework uses neural ordinary differential equations to learn the mapping between states and control inputs, allowing the system to learn unknown dynamics while simultaneously guaranteeing safety.
  2. Transformers: In the realm of heavy-duty transport, Jiaping Ren and Yuexin Ma developed Neural Predictive Control (NPC). They utilized a specialized Transformer architecture, NVFormer, to help autonomous trucks "feel" their own dynamics. By learning from historical fleet data, the system outperformed traditional physics models in fuel efficiency by over 3%, largely because it could account for complex engine behaviors that are too difficult to model by hand.

The Security Implications of Learned Dynamics

If a system's dynamics can be learned from observation for the purpose of control, they can also be learned for the purpose of subversion. André M. H. Teixeira and Daniel Arnström demonstrated a stealthy deactivation of safety filters by eavesdropping on a system's inputs and outputs. By reconstructing a topologically equivalent model of the system's observer dynamics in a latent space, an attacker can craft corrupted sensor data that tricks the system into a crash without ever knowing the physical parameters of the hardware.

Why It Works: The Power of Latent Representations

The success of these methods across drones, trucks, and power grids (Chen Shen) stems from a shared insight: dynamics are often simpler in the right representation.

Whether using Koopman operator theory to linearize power grid fluctuations or neural networks to map quadrotor settling times (Jiankang Zhao), these researchers are bypassing the "middleman" of human-derived physics. By letting the data define the model, they create controllers that are not only easier to build but are often more resilient to the messy, non-linear realities of the physical world.

Papers in this version

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_learning_dynamics_from_observations_for__347090b7,
  title = {Learning Dynamics from Observations for Black Box Control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-dynamics-from-observations-for-black-box-control/rdr_347090b7223949c584c125d4a9bee918},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_learning_dynamics_from_observations_for__347090b7_v1,
  title = {Learning Dynamics from Observations for Black Box Control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-dynamics-from-observations-for-black-box-control/rdr_347090b7223949c584c125d4a9bee918/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}