JALAN-Sim: A 200-Million-FPS Simulated Environment for 2D Navigation in Cluttered Spaces

Joshua Julian Damanik, Chala Adane Deresa, Sujeong Park, Wajih Imliki, and Han-Lim Choi

🛠️ This project is still in development. The paper and code will be available soon.

Abstract

JALAN-Sim, the Joint Autonomous LiDAR-Aided Navigation Simulator, is a massively parallel 2D robot navigation environment able to execute 200 million simulation steps per second on a single RTX 4090 GPU—more than 200 times faster than existing robotics simulators. The features include loading PGM occupancy maps, simulate LiDAR sensor using ray casting kernels, differential-drive state space, PID controller, and multi-point collision checks against static or key-framed dynamic obstacles. A gym-compatible Python bridge lets user train RL policy on most reinforcement learning libraries. The simulation is compatible with existing navigation datasets, including BARN and DynaBARN for navigation policy training and benchmarking in a cluttered spaces. The trained navigation policy with SAC hits 98% success tested with BARN dataset, and shown successful trials on sim-to-real experiment on indoor tracks with two differential-drive UGVs.

Experiments

Experiment 1: Training with BARN dataset

The trained policy using PPO finishes in under 2 minutes with 98% success rate validation on BARN dataset.

Experiment 2: Training with F1tenth map dataset

The trained policy using PPO finishes in under 2 minutes can navigate the whole lap without collision.

Experiment 3: Sim-to-Real

The trained policy using SAC can navigate the indoor track without collision at 0.7 m/s.