Control GPT

Control-AI Architect is an autonomous engineering agent that turns natural language into stabilized physics simulations. It solves the “Math-to-Code” bottleneck, allowing users to prototype complex dynamic systems in seconds without writing a single line of Python.

Here is the concrete user walkthrough. You start by typing a prompt like “Create an inverted pendulum on a moving cart” into the chat. The Agent immediately acts as a physicist: it derives the Equations of Motion (EOMs), generates the Python code, and calculates the system topology. You instantly see the result: a rendered Block Diagram of the system architecture and the raw LaTeX equations displayed on the screen.

Next, you click “Run Simulation.” Because the generated model is physically accurate, the pendulum falls over—the system is unstable by nature. This is where the innovation shines. You click “Start Auto-Design Loop.” You watch as the Agent writes a control algorithm, runs a simulation in the background, and analyzes the mathematical error. If the pendulum falls, the Agent detects the failure, adjusts its own control logic (e.g., tuning the PID gains), and retries.

Shutterstock

You see the progress bar update as it iterates until the physics allow the goal to be met. Finally, you click “Play Animation” and watch the pendulum balance perfectly upright in a real-time visualization.

The project demonstrates a true agentic workflow by using a physics engine as a ground-truth validator. It integrates OpenAI models with scipy for numerical integration and matplotlib for rendering, bridging the gap between generative text and deterministic engineering. To reproduce this, simply clone the repo, install the dependencies, and ask the Agent to build a “drone” or “robotic arm”—it will handle the derivation, coding, and stabilization end-to-end.

None, I am a research in control theory