Active Exploration

An agent deciding for itself where to look and what to touch, actively gathering information about the unknown.

As opposed to passively receiving data, active exploration means an agent chooses how to move and interact so as to gather information at as low a cost as possible. In navigation, this usually means placing a robot in an unfamiliar house and asking it to cover and map as much of it as possible within a limited number of steps, a prerequisite capability for downstream tasks like object search and embodied question answering. The classic approach is frontier exploration, heading toward the boundary between known and unknown areas; a learning-based method such as Active Neural SLAM (2020) builds a map with a neural network, with a global policy choosing a long-term goal and a local policy handling the walk there. In manipulation, active exploration can also mean pushing or shaking an object to discover hidden properties such as its mass or its articulated structure. “Exploration” in reinforcement learning is a related but broader concept.

ExampleIn the Habitat simulator, an Active Neural SLAM agent moves through an unfamiliar apartment on its own and draws a top-down map; a variant of the method also won the CVPR 2019 Habitat PointGoal Navigation Challenge.

Related

Frontier-based Exploration · Active Perception · Interactive Perception · Exploration vs. Exploitation · Simultaneous Localization and Mapping · Navigation

Sources

Learning to Explore using Active Neural SLAM (arXiv 2004.05155)