Robotic Arm Project Using Reinforcement Learning

Developed and trained a 7-DoF Fetch Mobile Manipulator with a two-finger parallel gripper to perform a block-pushing task using Reinforcement Learning. The project focused on applying and comparing multiple RL algorithms from Stable-Baselines3 (PPO, SAC, TD3) while experimenting with custom reward functions and hyperparameter tuning to optimize performance.

Key contributions included:

Designing and testing four different reward functions (dense, sparse, distance-based, time-penalty, and custom) to evaluate their impact on learning efficiency.

Implementing and comparing three RL algorithms, analyzing performance differences and identifying strengths of each approach.

Exploring hyperparameter variations and documenting their effect on training outcomes.

Writing clean, well-documented code with Gymnasium and Stable-Baselines3, and organizing results in a structured README/Notion report.

This project enhanced practical understanding of reinforcement learning, algorithm design, and performance evaluation while addressing a real-world robotic control problem.