Act to See
A closed-loop robotic system that interacts with an unknown articulated object and reconstructs an explicit URDF online, using the representation as both output and planner memory.
Read the projectI build AI that works
where life happens.
Selected work
All projects
OpenSUN3D affordance grounding
An Open-Vocabulary 3D Affordance Grounding system that combines 3D segmentation, language-guided proposals, and VLM verification to locate functional parts in 3D scenes.
Read the project
Real-time traffic anomaly detection
Tested the accuracy–latency trade-off of image-pretrained vision transformers, reaching 78.1% AUC-ROC at 200 FPS on DoTA.
Read the study
Human-in-the-loop 4D data engine
Built LiDAR–camera visualization and annotation systems used across perception teams, then trained a point-cloud sequence model to support dense pedestrian tracking workflows.
Project detailsHow I work
Product systems
I turn ambiguous needs into coherent native workflows, then carry them through privacy, reliability, distribution, and support.
On-device AI
I integrate speech, language, and vision models under real constraints: latency, memory, battery, offline use, and replaceable runtimes.
Perception & robotics
I build tools and models that connect sensor data to action — from production LiDAR pipelines to interactive 3D scene understanding.
Recent recognition
Get in touch