Work

MASLLO-Net: A Transformer-Guided Framework for Autonomous Mobility Aid Search Amidst Low Light and Occlusions

Year: 2026

Type: article

Abstract: Autonomous search for one or more objects in an unknown environment is of interest in applications such as service and assistive robotics. In practical scenarios, these robots must perform the tasks under non-ideal conditions, which include low illumination environments and partial occlusions. Further, these tasks have to be typically performed with constraints on sensing hardware. To address these challenges, we present a framework termed MASLLO-Net (short for Mobility Aid Search amidst Low Light and Occlusions Network) for autonomous search via a mobile platform equipped only with low-end vision and distance sensors. The proposed MASLLO-Net is a two-stage lightweight transformer-based framework which maintains robust perception under challenging conditions. MASLLO-Net is guided by the definition of new task-specific loss functions which handle low light and occlusions. In particular, we propose 1) an image enhancement-guided loss function that takes into consideration low-light images acquired by a vision sensor, and 2) an occlusion-aware loss function to account for partial scene observability as captured by the vision sensor. Experiments with an indigenously fabricated mobile robot incorporating the proposed framework are presented for the search of mobility aids in an unknown indoor environment.

Source: IEEE Access

DOI: 10.1109/access.2026.3735104

Authors: Nambala Ramsai, K. Suryanarayana, Pravin Nair, K. Sridharan

Institution: Indian Institute of Technology Madras

Language: English

Topic: Advanced Neural Network Applications

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

SDG: Industry, innovation and infrastructure

Open Access status: gold

Estimated APC paid: $2,075