Researchers from Samsung R&D Institute China-Beijing (SRC-B) won second place at the 2020 Conference on Computer Vision and Pattern Recognition (CVPR 2020) Embodied Artificial Intelligence (AI) Challenge. Embodied AI is an emerging field aiming to enable robots to understand human commands and perform correct actions within a virtual environment. The CVPR Embodied AI Challenge started in 2019, and it is now one of the most important events on embodied AI. This year, more than 10 world-class teams joined the contest. Placing second in this challenge has proven SRC-B’s outstanding capabilities in embodied AI technology and its strong aspiration to push the technology frontier.
△ Samsung Research China-Beijing (SRC-B) Team
The challenge consisted of two subtasks, point navigation and object navigation. The point navigation task involved general commands like “go 5 m north and 3 m east.” Meanwhile, the object navigation subtask included commands such as “find the TV,” which requires a robot to understand the command, explore the room, and avoid colliding with other objects to find the target.
SRC-B researchers chose to participate in the object navigation task because it is more closely related to future AI scenarios despite it being more challenging. Multiple modules were built to solve complex tasks such as SLAM, map reconstruction and optimization, path planning, and instance recognition and segmentation. In the end, SRC-B received a score of 0.0997, which is very close to that of the champion’s 0.1021.
As AI continues to be part of daily life, showing its irreplaceable power, its shortcomings also emerge: the lack of personalization, a weak connection to its environment, limited multimodality, heavy data, and costly training requirements. The idea of embodied AI is to let agents cognize, understand, learn, and act from interacting with the environment. The embodied AI is believed to be a next-generation AI paradigm because it will not only enable agents to have a better environmental and situational understanding but also reduce the data and graphics processing unit training loads. The ability of embodied AI can significantly improve the performance of future Internet of things (IoT), multidevice experiences, and robots.