SUTD and Southwest Jiaotong University Unveil AI Framework for Hierarchical Segmentation of 3D Spaces

On Sept. 30, 2026, techxplore.com reported on a joint study by Assistant Professor Na Zhao and other researchers at the Singapore University of Technology and Design (SUTD) and Southwest Jiaotong University, which developed a new AI framework for understanding 3D spaces at multiple levels.

The SUTD announcement highlights the need to translate 3D information from LiDAR and depth cameras into a level of detail that machines can use to act. The safety of robots and autonomous machines depends on how well they can distinguish doors, windows and chairs, rather than only broad categories such as furniture.

The report points to research moving toward machines distinguishing not only “what is there” in a 3D space, but also “at what level of detail.” ML3DHS focuses on 3D hierarchical semantic segmentation, assigning multiple related labels to the same point, from broad categories such as furniture to specific ones such as chairs and tables. What is new here is the effort to handle differences in granularity needed for action planning and safe contact decisions, rather than simple object recognition. At the same time, this is an extension of 3D recognition research, and appears to give concrete form to an existing trend of connecting sensor inputs with machine decisions.\n\nIn related coverage, SUTD reports control method for single-actuator flying robot, improving path tracking with predictive control examined a control method. This report focuses on hierarchical perception, which comes before control, showing how research at SUTD is advancing in parallel on “moving” and “understanding.” However, the article does not provide grounds to conclude that the two are directly connected or that implementation in autonomous mobility is imminent. Rather, the focus here is on the design of recognition labels, and any connection to control requires separate confirmation.\n\nFrom an industry perspective, the work concerns how robots use point clouds captured by LiDAR and depth cameras in real-world environments. Broad categories alone may not provide enough information to distinguish furniture from doors and windows, for example, when navigating narrow passages or deciding whether to make contact with an object. The use of hierarchical labels indicates that label systems and training design, not just sensor performance, can make a difference in implementation. The next points to examine are ML3DHS’s evaluation metrics, the datasets used, how it compares with existing methods and whether it has been tested on physical systems.

For Japan’s robotics and measurement-equipment developers, recognition systems that handle 3D point clouds hierarchically could become a competitive factor alongside sensor resolution. In applications such as service robots and indoor transport, distinguishing furniture, openings and obstacles at different levels of detail could make the development of training data and label systems an important issue.