The Stanford AI Lab (SAIL) is presenting a wide array of research at the Conference on Computer Vision and Pattern Recognition (CVPR) 2026, held in Denver, Colorado, from June 3 to June 7. The lab's contributions span diverse fields including video generation, robot learning, and medical imaging, with multiple papers receiving award nominations.

Notable research includes "BAgger," which introduces backwards aggregation to mitigate drift in autoregressive video diffusion models, and "BulletTime," a system for decoupled control of time and camera pose in 4D video synthesis. Other video-related works include "Choreographing a World of Dynamic Objects" for 4D motion generation and "Stand-In," a lightweight plugin for identity control in video models. "Generated Reality" explores interactive video generation with hand and camera control, while "GaussFusion" enhances 3D reconstruction using geometry-informed video generators and neural rendering (technique for generating photorealistic images from 3D data).

In the domain of robotics and embodied AI (systems that interact with a physical environment), "Ego-Pi" focuses on VLA (vision-language-action) fine-tuning for ego-centric human and robot data. "HoMMI" examines whole-body mobile manipulation through imitation learning, while "VULCAN" utilizes tool-augmented multi-agent systems for 3D object arrangement. Additionally, "Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment" has been named a Best Paper Finalist.

Other research highlights include "Physical Object Understanding with a Physically Controllable World Model" and "Spherical Leech Quantization," both receiving award nominations for their advancements in object understanding and visual tokenization (process of converting data into discrete units for processing) respectively. The lab is also presenting "GeoSAE," which uses sparse autoencoders (neural networks that learn compressed representations of data) to annotate brain MRI foundation models for clinical use in Alzheimer's disease. Finally, "Theory of Space" investigates how foundation models develop spatial beliefs through active exploration.

The Stanford AI Lab (SAIL) is presenting a wide array of research at the Conference on Computer Vision and Pattern Recognition (CVPR) 2026, held in Denver, Colorado, from June 3 to June 7. The lab's contributions span diverse fields including video generation, robot learning, and medical imaging, with multiple papers receiving award nominations.

Notable research includes "BAgger," which introduces backwards aggregation to mitigate drift in autoregressive video diffusion models, and "BulletTime," a system for decoupled control of time and camera pose in 4D video synthesis. Other video-related works include "Choreographing a World of Dynamic Objects" for 4D motion generation and "Stand-In," a lightweight plugin for identity control in video models. "Generated Reality" explores interactive video generation with hand and camera control, while "GaussFusion" enhances 3D reconstruction using geometry-informed video generators and neural rendering (technique for generating photorealistic images from 3D data).

In the domain of robotics and embodied AI (systems that interact with a physical environment), "Ego-Pi" focuses on VLA (vision-language-action) fine-tuning for ego-centric human and robot data. "HoMMI" examines whole-body mobile manipulation through imitation learning, while "VULCAN" utilizes tool-augmented multi-agent systems for 3D object arrangement. Additionally, "Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment" has been named a Best Paper Finalist.

Other research highlights include "Physical Object Understanding with a Physically Controllable World Model" and "Spherical Leech Quantization," both receiving award nominations for their advancements in object understanding and visual tokenization (process of converting data into discrete units for processing) respectively. The lab is also presenting "GeoSAE," which uses sparse autoencoders (neural networks that learn compressed representations of data) to annotate brain MRI foundation models for clinical use in Alzheimer's disease. Finally, "Theory of Space" investigates how foundation models develop spatial beliefs through active exploration.