# Group G25 - Formal & Title & Goals ## Team members - Diogo Costa - Johnny Fernandes - Nelson Neto ## (i) Title and General objectives **RL-Based Autonomous Shepherd Robot for Livestock Herding** - Implement effective herding behaviors through proximity and movement strategies - Build a 3D environment with realistic robot dynamics and LIDAR-based perception - Develop a mobile robot capable of autonomously guiding a flock of sheep into a designated target area using Reinforcement Learning # Group G25 - (ii) Intermediate Goals ## Intermediate goals - Set up the Webots simulation environment with an open field and target zone - Implement lightweight Gymnasium-based 2D herding environment - Design a Sheep and Dog robot - Implement a sheep flocking model for fast RL iteration - Validate LiDAR sensor feedback for sheep detection and distance estimation # Group G25 - Course Project (Final) Goals ## (iii) Main goals - State-of-the-art survey on shepherding algorithms and multi-agent RL herding - Train the robot using PPO to successfully herd a single sheep into the goal - Achieve fully autonomous herding of multiple sheep and a full flock into the target area - Optimize robot trajectory to minimize the time required to group the flock - Ensure zero collisions between the robot and the sheep during the task - Quantitative evaluation: success rate, time-to-pen, flock dispersion metrics - Article, demo video, and final presentation ## (iv) Extra Merit - Curriculum Learning (scaling from 1 sheep to a flock) - Comparison of performance between Differential Drive and Mecanum wheels - Robustness testing under sensor noise or varying sheep speeds, configurations and parameters - Multi-shepherd cooperative mode: 2 dogs learn role specialization (collector vs. driver) - Obstacles and terrain (walls, narrow gates, corridors) the flock must be funneled through # Group G25 - Tools & Limitations ## (v) Tools - Webots for 3D physics simulation with ROS2 integration via `webots_ros2` package - Stable-Baselines3 for the PPO algorithm implementation - Gymnasium (OpenAI) for the RL environment wrapper (lightweight 2D herding env for fast RL training) - Python as the primary programming language (sheep flocking model, reward shaping, evaluation) ## (vi) Limitations - Computational Power: Training time might be high for complex flock behaviors - Sim-to-Real Gap: No real-world validation of the herding controller; project is simulation-only (2D + Webots 3D) - Model Complexity: Simplified sheep behavior (scripted) may not account for all biological livestock nuances