Abstract
Recently, learning-based approaches show promising results in navigation tasks. However, the poor generalization capability and the simulation-reality gap prevent a wide range of applications. We consider the problem of improving the generalization of mobile robots and achieving sim-to-real transfer for navigation skills. To that end, we propose a cross-modal fusion method and a knowledge transfer framework for better generalization. This is realized by a teacher-student distillation architecture. The teacher learns a discriminative representation and the near-perfect policy in an ideal environment. By imitating the behavior and representation of the teacher, the student is able to align the features from noisy multi-modal input and reduce the influence of variations on navigation policy. We evaluate our method in simulated and real-world environments. Experiments show that our method outperforms the baselines by a large margin and achieves robust navigation performance with varying working conditions.
Architecture
To improve the generalization ability of a learning-based navigation approach for sim-to-real transfer, we use a distillation architecture instead of directly using the RL+Domain Randomization, because domain randomization will increase the learning difficulty for reinforcement learning. The teacher module is a pure-RL approach trained in a perfect environment (without noise). The student module is trained in a corresponding noisy environment by imitating the behavior of the teacher in perfect environment. Contrastive learning is introduced here to ensure the representation network can be sensitive to the noise signal.
Experiments
We evaluate our methods in different speed settings, noise settings, and scene settings, our distill architecture outperforms the RL + Domain Randomization by a large margin.
Summary Video
Citation
@article{Cai2023RobustNW,
title={Robust Navigation with Cross-Modal Fusion and Knowledge Transfer},
author={Wenzhe Cai and Guangran Cheng and Lingyue Kong and Lu Dong and Changyin Sun},
journal={2023 IEEE International Conference on Robotics and Automation (ICRA)},
year={2023},
pages={10233-10239}
}