More about HKUST
A Survey on Open-World Spatial Perception and Reasoning
PhD Qualifying Examination
Title: "A Survey on Open-World Spatial Perception and Reasoning"
by
Mr. Yang CAO
Abstract:
Open-World spatial perception and reasoning is emerging as a core capability
for AR, VR, and embodied AI, where an agent needs to localize, recognize, and
reason about objects in 3D space beyond fixed categories, sensors, or query
types. Riding on foundation models such as CLIP, VGGT, and multimodal LLMs,
recent research has advanced 3D spatial understanding along three
complementary axes. A vocabulary axis pursues open-vocabulary 3D object
detection beyond a fixed category set; a sensor axis moves toward
sensor-geometry-free detection that reduces reliance on calibrated multi-view
poses or depth, enabling broader deployment; and a semantic axis targets
fine-grained spatial reasoning over relational and compositional queries.
Together they trace a progression from closed to open perception and from
perception to cognition. We first formalize the tasks and their core
challenges: category openness, minimal sensor assumptions, and generalization
to complex queries. We then organize detection methods by input modality
(multi-view RGB, point clouds) and degree of vocabulary openness, categorize
spatial reasoning by modeling framework and training strategy, and introduce
datasets and evaluation protocols alongside the methods. We aim to provide a
common foundation to catalyze advances in open-world spatial perception and
reasoning.
Date: Wednesday, 2 September 2026
Time: 10:00 - 12:00
Venue: Room 3494
Lift 25/26
Committee Members: Dr. Dan Xu (Supervisor)
Dr. Qifeng Chen (Chairperson)
Dr. Chaojian Li