all projects
Deep RL · Gazebo

navbot_ppo

Mapless mobile-robot navigation: deep-RL motion planners — PPO, DDPG, SAC and TD3 — that drive a TurtleBot to a goal from sparse LiDAR and the target pose, with no SLAM and no map.

Result PPO · DDPG · SAC · TD3 on one 16-D LiDAR state Stars ★ 55 Tags PPO · DDPG · SAC · TD3 · ROS · LiDAR
View on GitHub Paper · arXiv:2405.16266

A reinforcement-learning motion planner that learns to navigate from scratch. The agent’s state is a compact 16-D vector — downsampled LiDAR ranges plus the relative goal — and its action is continuous wheel velocity. There is no global map and no classical planner in the loop: the policy itself is the planner.

Highlights

  • Mapless. Navigation from sparse LiDAR + target pose only — robust to unseen layouts.
  • Four deep-RL algorithms. PPO, DDPG, SAC and TD3 implementations (PyTorch), all trained on the same 16-D LiDAR observation so the continuous-control policies can be compared head-to-head.
  • Reproducible. Dockerized Gazebo GUI so the simulation comes up with one command.
  • Real-time avoidance. The overhead demo shows the agent sweeping LiDAR, weaving around obstacles, and driving to the goal.

This repository is the practical, runnable counterpart to my publication on safe mobile-robot navigation with enhanced PPO.