Zero to Autonomy in Real-Time image

Zero to Autonomy in Real-Time

Project Overview

How can robots adapt to terrains that evolve over time? We develop a way for robots to update their dynamics models on-the-fly as their terrain changes during autonomous missions.

Skills Used

Python ROS (Noetic) PyTorch Unity Neural Networks Model Predictive Control Dynamics Modeling

Motivation

Sudden dynamics shifts, like moving from pavement to ice, can destablilize robot control unless the model updates in real time. Our system adapts to terrain changes online within seconds, maintaining MPC accuracy, even on extremely slippery terrain.

Overall Framework

figure1.png

Model adapatation and model-predictive control occur simultaneously.

  1. Adaptation: Our method updates the robot’s dynamics model in real time using live sensor data.
  2. Control: MPPI controls the robot using trajectory rollouts generated by the latest adapted model.

Results

ice_results.gif

Technical Contributions

  • Developed real-time coefficient updates for neural network dynamics models.
  • Planned and directed robot autonomy experiments at a local ice-skating rink.
  • Integrated neural network dynamics models with MPPI control on-board hardware robots.

Publication (Under Review)

W. Ward, S. Etter, J. Quattrociocchi, C. Ellis, A. J. Thorpe, U. Topcu, “Zero to Autonomy in Real-Time: Online Adaptation of Dynamics in Unstructured Environments,” Under review for ICRA 2026.

Additional Details