Navigation is a key task in building an intelligent agent. Regardless of the task, the core problem for navigation in an unknown environment is exploration, which means how to effectively access as many environments as possible. This is useful for maximizing coverage to provide the best opportunities for finding targets in unknown environments or to effectively pre-map environments in limited time budgets.
The new paper Learning to Explore Using Active Neural Slam from researchers at Carnegie Mellon University, Facebook AI Research, and University of Illinois at Urbana-Champaign, introduces Active Neural SLAM, a modular and hierarchical approach to learning policies for exploring 3D environments.
Previous research has used an end-to-end learning approach to solve this problem. But such end-to-end learning work for exploration relies on the use of imitation learning and millions of experience frameworks. It can be excessively expensive and performance is still no better than traditional methods that require no training at all.
The Active Neural SLAM approach benefits from the strengths of both classical and learning-based methods by using analytical path planners with learned SLAM modules, and global and local policies. This enables the method to retain advantages that learning has to offer while eliminating the problems associated with full-blown end-to-end learning.
The proposed Active Neural SLAM exploration architecture is composed of a learned Neural SLAM module, a global policy, and a local policy, which are interfaced through a map and an analytical path planner. The learned Neural SLAM module generates free space maps and estimates agent pose based on the input RGB images and motion sensors. These maps and poses are used by the global policy to formulate a long-term goal, which is converted to a short-term goal for the local policy using an analytic path planner. This local policy is trained to navigate to this short-term goal by using learning to directly map original RGB images to the operation that the agent should perform.
The Active Neural SLAM model and all baselines of the “Exploration” mission for 10 million frames were trained on the Habitat simulator with the Gibson and Matterport (MP3D) datasets. The proposed model achieved an average absolute and relative coverage of 32.701m2/0.948 as compared to 24.863m2/0.789 for the best baseline. This demonstrates the Active Neural SLAM model is more effective in exhaustive exploration compared to the baseline. The improvement is attributed to a reduction in the horizon of the long-term exploration problem with hierarchical strategy architecture, with the global policies instead taking tens of low-level actions in terms of navigation.
PyTorch implementation of Active Neural SLAM and pretrained models are available on the project’s GitHub. The paper Learning to Explore Using Active Neural Slam is on arXiv.
Author: Xuehan Wang | Editor: Michael Sarazen