🧠🤖 Cognitive Architecture for an Autonomous Robot in Robobo Sim

🌟 Project Overview

This project is part of the Intelligent Robotics II course for the academic year 2023/2024. We focused on designing and implementing a Cognitive Architecture for an autonomous robot in the Robobo Simulator, using Python and some awesome libraries! 🐍

Authors

• Ettore Caputo
• Stefano Iannicelli

🚀 Introduction

Our goal was to build a smart robot that could navigate autonomously and make decisions on its own in a simulated world 🌍🛣️. We designed a deliberative decision-making system that includes a World Model and a Utility Model.

🤔 Problem Description

How can a robot understand the world around it, predict what will happen next, and choose the best thing to do — all on its own? That's the challenge we tackled! 🔍⚙️

🕹️ Action Execution

The robot moves using its gyroscope sensor 🔄. Rotations are calculated with specific math formulas to guide its motion precisely. 📐🤖

🧩 Proposed Solution

🌐 World Model

This part predicts the robot’s future state based on its current situation and actions. It comes in two flavors:

🔢 Mathematical Version

• Uses good ol’ algebra to estimate future positions
• Integrates sensor data to update its understanding of the world

🧠 Neural Network Version

• Includes two neural networks:
  • One for position prediction
  • One for sensor data prediction
• Learns to generalize the robot’s behavior from experience

🎯 Utility Model

Helps the robot decide what to do next by evaluating how "good" different future states are.

🔍 Intrinsic Module

• Encourages the robot to explore new things using a novelty function
• Tracks past experiences to measure how “new” something is

🏆 Extrinsic Module

• Uses a neural network to evaluate predicted states
• Trained with Binary Cross-Entropy loss
• Rewards states that bring the robot closer to its goal

🧪 Training Strategy

• Follows an ϵ-greedy policy for choosing actions
• Starts off curious (focused on novelty)
• Gradually learns from experience using the neural network
• With the mathematical model, it can learn a good policy in just ~10 minutes! ⏱

🛠️ Challenges and Solutions

• The neural World Model had trouble accurately predicting future states
• One issue was likely incorrect object size perception
• This led to strange behavior — the robot avoided obstacles but didn’t go anywhere
• Future improvements should aim to refine object perception and improve model accuracy

✅ Conclusions

• We built a simple but effective cognitive architecture
• The mathematical World Model enabled the robot to learn purposeful behavior after a short exploration phase
• The neural version shows promise but needs more training and tuning

🔮 Future Work

• Improve the neural World Model
• Try out different reward functions to fine-tune learning
• Test in more complex environments to push the limits

---

📄 For more details, check out the project report! Thanks for reading! 😊🚀