Two Papers Accepted to NeurIPS 2026: TAR and Physical AI Smart Spaces

I am excited to share that two of our papers have been accepted to the NeurIPS 2026 Evaluations and Datasets Track. Both projects build benchmarks for physical AI systems that must understand complex activity across real-world infrastructure, moving beyond isolated perception tasks toward richer reasoning and 3D scene understanding.

From Detection to Understanding – A Multi-Task Dataset for Traffic Anomaly Reasoning (TAR) introduces a large-scale resource for training and evaluating vision-language models on traffic anomaly understanding. TAR provides 44,040 annotations with explicit reasoning traces across 10 task types for 3,670 CCTV transportation videos. The tasks progress from question answering to temporal reasoning and scene understanding, helping reveal where strong detection performance does not yet translate into genuine event comprehension. TAR also serves as the official training data for AI City Challenge 2026 Track 3.

Physical AI Smart Spaces: A Large-Scale Benchmark for Multi-Camera 3D Perception in Smart Spaces (PAISS) brings together more than 280 hours of synchronized 1080p video from nearly 1,800 cameras across warehouses, hospitals, retail venues, and related environments. Its pipeline spans Isaac Sim synthetic generation, Cosmos Transfer appearance augmentation, and real-world Sim2Real evaluation, with annotations and protocols for multi-camera identities, 2D and 3D boxes, calibration, depth, and 3D tracking accuracy. The PhysicalAI-SmartSpaces dataset has now received over 800,000 downloads, reflecting strong community interest in scalable multi-camera perception for physical AI.

I am grateful to all of our collaborators whose work made these datasets, evaluation systems, and papers possible. Together, TAR and PAISS support a broader shift toward physical AI benchmarks that measure not only what systems detect, but also how well they reason over events, geometry, time, and interactions across complex environments.