Kunyu Peng

PROVIA ranks mistakes best among the evaluated controlled baselines on CaptainCook4D, IndustReal, HoloAssist and IMPACT-ego, under a validation false-alarm budget and against controls that use timing alone.

INSPECT is introduced, which learns robot view preferences from records of a smart-glasses assistant that answers part queries and provides next-step guidance and achieves the highest view utility among the compared non-oracle policies and raises human-rated full verifiability.

CoRef-GS is proposed, a cooperative referring Gaussian splatting framework that constructs local open-vocabulary instance-aware Gaussian maps, then aligns partially overlapping maps with a cross-agent alignment module by geometric and semantic consistency, and grounds queries using a view-conditioned mask relation graph.

Results show that the multi-agent design of RoboFind fits the demands of personalized object search, where verifying object identity before declaring completion is what makes the outcome something a user can rely on.

Parallel execution of tasks in assistive robots can cut mission completion time by over 40%, revolutionizing how we think about robot coordination in dynamic environments.

X$^2$Localizer boosts single-frame retrieval performance by over 4% while maintaining full-video accuracy, bridging the gap between evaluation benchmarks and real-world applications.

Over 1,500 submissions revealed stark differences in model performance across diverse domains, highlighting the challenges of generalizing egocentric video understanding.

HGeo-TopoMap achieves remarkable gains in topological mapping accuracy and robustness, tackling the elusive challenge of centerline detection in autonomous driving.

Transforming point seeds into temporally consistent object instances, PS-Track sets a new benchmark in multi-object tracking without relying on traditional bounding box annotations.

HOWTransfer achieves an 86% success rate in translating human hand motions into robot actions, surpassing teleoperation in user preference.

Correcting errors in long-video understanding doesn't have to be a nightmare: IMPACT-CYCLE slashes human arbitration costs by 4.8x while boosting VQA accuracy by intelligently decomposing the task and focusing human effort where it matters most.

Autonomous vehicles can now better identify the unexpected, thanks to a new method that boosts out-of-distribution detection by up to 20% without retraining.

Achieve state-of-the-art panoramic segmentation by training on local perspective views and generalizing to full 360° images, even with geometric distortions and unseen classes.