6.4210 Project. By Aileen Liao and Daniel Prakah-Asante
"Robotic Soccer Goalie" is a research initiative aimed at evaluating various learning and physics-based controllers for dynamic tasks, focusing on a robotic goalkeeper's ability to intercept soccer shots in simulated environments.
- Simulation of a dynamic soccer environment using the Drake framework.
- Implementation of diverse controllers: physics-based, neural network-based, and nueral ode.
- Data collection from simulations to train and evaluate machine learning models.
Run the simulation script main.py notebook and modify controllers to test different controllers.
- Physics-Based: Utilizing classical mechanics for trajectory predictions.
- Sampling-Based: Exploring various trajectories under different spin conditions.
- Neural Networks:
- RNN: Predicting blocking positions based on the last three ball positions.
- GRU: Using gated recurrent unit cells to predict blocking positions based on the last three ball positions.
- ODE-LSTM: Combining LSTM with Neural ODEs for dynamic predictions.
- Controllers are evaluated over 50 simulation episodes.
- Metrics: Success rate in blocking shots under various conditions.
- Physics-based approaches demonstrated superior performance in dynamic tasks.
- Neural ODEs showed promise for environments with unknown dynamics.
- Investigating Continuous Neural Networks for varied dynamic environments.
- Real-world controller evaluations for practical applicability.
Special thanks to Professor Russell Tedrake, Ethan Yang, and the MIT 6.4210 staff.
This project is distributed under the MIT License.
If you use or reference this project, please cite as follows:
@misc{robotic_soccer_goalie,
title={Robotic Soccer Goalie: Evaluation of Learning and Physics-Based Controllers for Dynamic Tasks},
author={Prakah-Asante, Daniel and Liao, Aileen},
year={2023},
publisher={GitHub},
journal={GitHub repository},
howpublished={\url{https://github.com/username/robotic-soccer-goalie}}
}