LiCS: Navigation using Learned-imitation on Cluttered Space

TL;DR

LiCS tackles robust local navigation for UGVs in cluttered indoors by learning to imitate an optimization-based planner through a Transformer-based policy trained offline with Gaussian noise. A safety check layer provides real-time geometric collision avoidance, enabling fast operation up to $1.5\ ext{\,m/s}$ while navigating narrow passages. The approach is validated in Gazebo simulations and on a Jackal platform, showing superior performance to baselines and strong generalization to unseen clutter, albeit with limitations on highly constrained tracks. The work suggests that combining offline imitation with a lightweight Transformer and a principled safety layer can yield practical, robust navigation for real-world UGVs, with future directions in integrating global planning and SLAM.

Abstract

In this letter, we propose a robust and fast navigation system in a narrow indoor environment for UGV (Unmanned Ground Vehicle) using 2D LiDAR and odometry. We used behavior cloning with Transformer neural network to learn the optimization-based baseline algorithm. We inject Gaussian noise during expert demonstration to increase the robustness of learned policy. We evaluate the performance of LiCS using both simulation and hardware experiments. It outperforms all other baselines in terms of navigation performance and can maintain its robust performance even on highly cluttered environments. During the hardware experiments, LiCS can maintain safe navigation at maximum speed of $1.5\ m/s$.

Figures (6)

• Figure 1: Demonstrations with exploration noise allow policy to learn how to effectively act on various states
• Figure 2: Training pipeline diagram of the Learned-imitation on Cluttered Space model. It consists of two steps: dataset acquisition (right) and offline learning (left). During dataset acquisition, observation data and optimal action given by baseline local planner are recorded. During the training, the optimal action and predicted action by the network are compared and MSE Loss is calculated for back-propagation.
• Figure 3: ROI illustration for safety check layer during linear and radial movement.
• Figure 4: Comparison results of local navigation performance of baseline algorithms with static global planner and various maximum velocities. In this results, LiCS was implemented without safety check layer on unseen environments (benchmark worlds)
• Figure 5: (a) Average scores of algorithms for each benchmark worlds group. (b) Average traversal times across benchmark worlds.