Back to workReinforcement Learning
Robot Navigation 2D
Custom 2D LiDAR robotics environment with scratch DQN and PPO navigation agents.
The problem
Mobile robot obstacle avoidance requires real-time decision making under continuous sensory inputs, dynamic moving obstacles, and sparse goal-reaching rewards.
Approach
Engineered a modular Gymnasium simulation featuring dynamic obstacles and a vectorized 200-ray LiDAR sensor. Implemented custom Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) agents from scratch with reward shaping, evaluation harnesses, and automated video recording.
At a glance
DQN & PPO
Algorithms
Implemented from scratch
200-Ray LiDAR
Sensor Model
Vectorized raycasting
Custom Gymnasium
Environment
Dynamic obstacle simulation
Headless & Visual
Evaluation
MP4 rollout generation
Stack
PythonPyTorchGymnasiumReinforcement LearningPyGameNumPy