Robot Navigation in Initially Unknown Environments using Manifold Alignment

We consider the problem of robot path planning in initially unknown environments using machine learning techniques. Previous research on this topic abstracts the similarity between navigation tasks in terms of a reward received by the robot and prescribes techniques to select the robot’s actions based on a policy that is learned through reusing previous policies. However, such techniques are susceptible to the geometry of environment features (e.g., obstacles), and do not ‘transfer’ well to new, different environments or domains. To address this problem, we propose a new algorithm based on the concept of manifold alignment from text mining. Robots first build a library of significant environment features from a source domain. Then, while navigating in a target domain they dynamically learn a mapping function between features across the two domains and use it to map features perceived in the target domain back to the source domain, and, probabilistically prescribe a commensurate action from the source domain. In order to verify the proposed technique, we have tested it in different indoor environments constructed in a test arena using a single Corobot robot