Neural-Network Based Adaptive Proxemics-Costmap for Human-Aware Autonomous Robot Navigation

摘要

Over the years, with the emergence of robots into the manufacturing industry, the manufacturing efficiency have increased. The continuous advancement of robot technologies in the field of navigation including localization, path planning, costmap design, obstacle sensing and tracking have drove revolutions to the industry. In the recent effort of transforming the industry towards Industry 4.0 [1], robot navigation remains a crucial part for a robot to co-exist with human co-workers. Humans prefer to maintain a certain proxemic distance [2] between themselves and robots for safety Abstract: In the revolution of Industry 4.0, autonomous robot navigation plays a vital role in ensuring intelligent cooperation with human workers to increase manufacturing efficiency. Human prefers to maintain a proxemic distance with other subjects for safety and comfort purposes, where the human personal-space can be represented by a costmap. Current proxemic costmaps perform well in defining the proxemic boundary to maintain the human-robot proxemic distance. However, these approaches generate static costmaps that are not adaptive towards different human states (linear position, angular position and velocity). This problem impacts the robot navigation efficiency, reduces human safety and comfort as the autonomous robot failed to prioritize avoiding certain humans over the other. To overcome this drawback, this paper proposed a neural-network based adaptive proxemic-costmap, named as NNPC, that can generate different sized personal-spaces at different human state encounters. The proposed proxemic-costmap was developed by learning a neural-network model using real human state data. A total of three human scenarios were used for data collection. The data were collected by tracking the humans in video recordings. After the model was trained, the proposed NNPC costmap was evaluated against two other state-of-art proxemic costmaps in five simulated human scenarios with various human states. Results show that NNPC outperformed the compared costmaps by ensuring human-aware robot manoeuvres that have higher robot efficiency and increased human safety and comfort.