Kernkonzepte
Pretraining a robust visual navigation system using BEV representations for Zero-shot Sim2Real transfer.
Zusammenfassung
I. Abstract
Introducing a novel system for Zero-Shot transfer from simulator to real world using BEV images.
Incorporating state-checking modules and LSTM for robustness in real-world scenarios.
II. Introduction
Sim2real gap challenges in transferring models to real-world applications.
Traditional methods vs. recent works like Style Transfer and Domain Adaptation.
III. Proposed Method
Perception model with ResNet-50 for compact representations compatible with downstream policy learning.
Enhancing robustness with Temporal State Checking (TSC) and Anchor State Checking (ASC).
IV. Experimental Platform and Setup
Hardware setup with Non-Holonomic, Differential-drive robot Beobotv3.
Utilizing Coral EdgeTPU for inference through ROS middleware.
V. Evaluation and Results
Policy learning experiments using PPO algorithm in CARLA simulator.
Planning experiments using TEB planner with BEV reconstructions.
VI. Discussion and Future Work
Decoupling perception model from control model for pretraining encoder.
Potential use of additional datasets for further training.
Statistiken
1 The authors are with Thomas Lord Department of Computer Science, University of Southern California, 90089, USA Correspondence to klekkala@usc.edu