News

INSAIT announces 17 papers at CVPR 2026, its strongest computer-vision presence so far. The accepted work spans robotics, spatial reasoning, egocentric vision, earth observation, visual agents, and agentic vision.

Nikolay Nikolov presents INSAIT’s latest Physical AI work at Google Cloud Day Sofia, focusing on systems that scale beyond robots trained only from demonstrations.

Bulgarian government ministers visit INSAIT and meet the Computer Vision group leadership, including Luc Van Gool, to discuss the institute’s next stage of development and strategic role for Bulgaria.

Asen Nachkov discusses autonomous-vehicle research and DiffSim Trinity in Bulgarian media. The work studies differentiable simulation as a path toward safer, more data-efficient autonomous driving.

Lorenzo Venturoli joins INSAIT for a Master’s thesis in 3D computer vision, bringing experience in 3D semantics, SLAM, and multimodal benchmarks.

EgoNight is released at ICLR 2026 as a benchmark for nighttime egocentric vision. The project introduces aligned day-night videos, nighttime VQA, depth estimation, and cross-illumination retrieval to test vision systems beyond daytime conditions.

Anna-Maria Halacheva discusses Articulate3D, robotics, and her research path on the Superhuman podcast with Georgi Nenov.

VOID is launched in collaboration with Netflix as an open model for video object and interaction deletion. The model removes objects while reconstructing physically plausible scene dynamics rather than only filling pixels.

Hongyu An joins INSAIT as a PhD student in Computer Vision with Luc Van Gool and Jinjin Gu. His work focuses on image/video super-resolution, restoration, and generative diffusion models.

Mengshun Hu joins INSAIT as a postdoctoral researcher in Computer Vision. His research focuses on video restoration, space-time video super-resolution, and emerging VideoAgent systems.

INSAIT announces 8 accepted ICLR 2026 papers, placing Bulgaria first in Eastern Europe by accepted publications at the conference. Several of the accepted works connect to vision, multimodal AI, and embodied reasoning.

Taewoo Kim joins INSAIT as a postdoctoral researcher in Computer Vision and computational imaging. His work connects event-based cameras, low-light enhancement, video generation/editing, and agentic AI.

INSAIT launches the CVPR 2026 Workshop on Agentic AI for Visual Media, led by Jinjin Gu and Lei Sun in collaboration with Snap and Adobe. The workshop focuses on AI systems that reason, plan, and execute multi-step visual-media workflows.

Dannong Xu joins INSAIT as a PhD student in Computer Vision, working on multimodal learning, large language models, and visual reasoning.

Asen Nachkov discusses DiffSim Trinity and safer autonomous driving on Bloomberg TV Bulgaria, explaining how differentiable simulation supports more reliable planning.

Asen Nachkov explains autonomous vehicles and real-world decision-making on Bulgarian National Radio, highlighting the role of simulation and prediction in driving systems.

Ruibo Ming joins INSAIT as a PhD student in the Computer Vision group, working on video generation, autonomous driving, 3D reconstruction, and multimodal systems.

Anna-Maria Halacheva discusses INSAIT’s CSRankings success and AI talent development on bTV, following the institute’s rise in European computer-science rankings.

INSAIT reaches 13th in Europe and 1st in Eastern Europe in CSRankings, with computer vision ranked 6th in Europe among the institute’s tracked research areas.

Nikola Popovic presents INSAIT’s large-scale 3D dataset work in a Bulgarian National Radio interview, explaining why open 3D data matters for language-aware scene understanding.

INSAIT releases SceneSplat-49k, the largest open-source 3D Gaussian Splatting scene collection, together with a benchmark for language-aware 3D AI systems.

INSAIT attends NeurIPS 2025 with 7 accepted papers and multiple workshop activities, including work connected to world models, video AI, and interactive environments.

Prof. Angela Yao visits INSAIT and meets the Computer Vision group, sharing perspectives from her work in human-centric and visual AI research.

StateSpaceDiffuser is accepted to NeurIPS 2025 as a diffusion-based world model for long-context spatial consistency. The work is led by Nedko Savov with collaborators including Deheng Zhang, Danda Paudel, and Luc Van Gool.

Cheng Tang joins INSAIT as a Computer Vision intern with Luc Van Gool, focusing on RL-driven multimodal reasoning enhancement.

Letian Shi joins INSAIT as a doctoral student in the Computer Vision group, working with Luc Van Gool and Danda Paudel.

INSAIT presents 3 accepted robotics papers at IROS 2025, including work on autonomous-vehicle policies trained with differentiable simulation.

Nikolay Nikolov presents SPEAR-1 and robotic foundation models on bTV, describing how 3D understanding can improve robot behavior in the physical world.

SPEAR-1 is featured on Bulgarian National Television as a robotics foundation model built around spatial reasoning and 3D understanding.

Jan-Nico Zaech gives an ICCV 2025 workshop keynote on robotic AI and industrial transfer, highlighting how foundation data can support more capable robots.

Stefan Ailuro joins INSAIT as a PhD student in Computer Vision, working with Luc Van Gool and Danda Paudel.

INSAIT announces 13 ICCV 2025 papers and a major computer-vision presence in Hawaii, including work across 3D vision, robotics, perception, and multimodal AI.

INSAIT co-organizes the OpenSUN3D challenges at ICCV 2025 with Stanford, ETH Zurich, Google, NVIDIA, Microsoft, and TUM, advancing open-world 3D scene understanding.

INSAIT’s robotics team presents at CoRL 2025 and qualifies for the Real Robot Challenge, bringing the group’s embodied-AI work into a leading robotics venue.

Generalist Robot Manipulation Beyond Action Labeled Data is accepted to CoRL 2025. The work studies how robot policies can learn from broader visual and 3D data, not only action-labeled demonstrations.

GAIA 2025 takes place in Sofia, bringing together researchers and industry around geospatial AI, foundation models, and Earth observation.

INSAIT announces GAIA 2025, the first international symposium on Geospatial AI and Applications with Foundation Models, hosted in Sofia.

Xiaoye Wang joins INSAIT for a long-term research visit in 3D computer vision, working with Luc Van Gool and Danda Paudel.

INSAIT opens Google co-funded PhD positions in egocentric vision and multimodal LLMs, supporting research on first-person perception, streaming multimodal models, and real-world visual understanding.

Yuanqi Yao joins INSAIT as a PhD student in Computer Vision and Robotics, focusing on embodied AI, robotic manipulation, and vision-language-action models.

Google gives INSAIT $150,000 to support multimodal LLM research under Luc Van Gool, extending the institute’s work in vision-language and egocentric AI.

Berke Gokmen joins INSAIT for a long-term research visit in Computer Vision and AI, working on 3D scene generation and generative vision models.

INSAIT co-organizes the OpenSUN3D Workshop at ICCV 2025, focused on open-vocabulary and open-world 3D scene understanding.

INSAIT co-organizes the Physics-IQ Challenge at ICCV 2025, a benchmark for evaluating whether generative models produce physically plausible videos.

Prof. Richard Hartley visits INSAIT and meets the Computer Vision group, giving a talk on modeling probability distributions in data manifolds using diffusion.

INSAIT co-organizes the NeurIPS 2025 Workshop on Embodied World Models for Decision-Making, connecting simulation, video-language-action models, robotics, and autonomous driving.

GaussianVLM is released as a 3D vision-language model for understanding immersive Gaussian-splat scenes reconstructed from ordinary video.

Zhendong Li joins INSAIT as a doctoral student in Computer Vision, working with Luc Van Gool and Danda Paudel.

INSAIT welcomes the 2025 AI Summer Research Fellows, selected from more than 4,000 applicants, with projects spanning computer vision, robotics, foundational models, and trustworthy AI.

INSAIT starts a joint research program with MIT CSAIL, expanding collaboration across frontier AI, systems, and related research directions.

ObjectRelator wins 2nd place in the Ego-Exo4D Challenge at the CVPR 2025 EgoVis Workshop, advancing cross-view object relation understanding for egocentric and exocentric video.

INSAIT’s work on multi-modal semantic segmentation under sensor failures wins Best Paper at the CVPR 2025 TMM-OpenWorld Workshop. The paper studies robustness when sensors are missing or noisy in real-world multimodal systems.

INSAIT presents 7 accepted papers at CVPR 2025, strengthening Bulgaria’s presence at the leading computer-vision conference.

SceneSplat is released for instant 3D object discovery from text in photorealistic Gaussian-splat scenes. The system enables open-vocabulary interaction directly in 3D space.

EarthMind is released as a foundation model for multimodal Earth observation and geospatial AI, combining visual, radar, and multi-spectral satellite data.

Exploration-Driven Generative Interactive Environments is accepted to CVPR 2025. The work introduces an auto-exploration pipeline for training generative interactive environments without human demonstrations.