Highlights and takeaways from the AI for Robotics workshop

This year’s workshop, with the special theme of Spatial AI ,brought together researchers working at the intersection of perception, geometry, learning and control to confront a shared challenge of how to build systems that can robustly perceive, reason about and act in the physical world. Across robotics, AR and autonomous driving, a clear picture emerged of a rapidly advancing field grappling with scale, dynamics and deployment in real environments. Several recurring themes cut across the talks and you can view them by clicking on the speaker photo.

SLAM, systems and the limits of scale

Daniel Cremers

Andrew Davison

While learning-based approaches dominated many discussions, the workshop repeatedly returned to the sobering reality that systems matter. Andrew Davison (Imperial College London) and Daniel Cremers (Technical University Munich) reflected on decades of progress in (V)SLAM, emphasizing that future breakthroughs will not come from algorithms alone. Efficient pipelines, tight integration between software and hardware and energy-aware computation will be decisive especially as the reliance on ever-larger data centers becomes untenable for embodied systems.

Marc Pollefeys

This systems view was complemented by reflections on data and compute. Marc Pollefeys (ETH Zurich and Microsoft) and Cordelia Schmid (Inria and Google) underscored the field’s growing dependence on large, high-quality datasets, while acknowledging a widening gap between academia and industry. Pollefeys argued that while frontier-scale models will remain largely in the hands of big tech due to GPU demands, academic research must remain closely coupled to industrial advances to stay relevant and impactful.

Learning at scale: datasets, world models and foundation models

Aniruddha Kembhavi

Fatma Güney

Across application domains, speakers highlighted a shift toward large-scale learning paradigms. In autonomous driving, Aniruddha Kembhavi (Wayve) presented world models that capture not only geometry but also dynamics and semantics, enabling large-scale training and evaluation of driving policies. Fatma Güney (Koç University) addressed the tension between model capacity and real-time constraints, introducing dual-system architectures that combine fast, reactive models with slower, more deliberative ones allowing systems to “think fast and slow” depending on context. Her work on counterfactual reward learning further demonstrated how aligning reward prediction with driving performance can significantly improve generalization.

Cordelia Schmid

In robotics, Cordelia Schmid explored vision-guided manipulation through the lens of foundation models. She showed how large language models can guide 3D manipulation policies, introduced GEMBench as a benchmark for such systems and presented Guardian, a tool for detecting planning and execution errors. A recurring insight was the untapped value of human video data: everyday videos of people performing tasks can serve as a rich supervisory signal for learning robot policies and grounded video captions.

Donghwan Lee

Donghwan Lee showcased NAVER LABS’ long-term investment in Physical AI, from early robotics work to recent foundation models such as DUSt3R. He demonstrated how these models enable applications ranging from digital twins to city-scale visual localization and how research is transitioning into deployed services.

Robots amongst humans: social interaction and evaluation in the wild

Michael Gienger

A distinct thread of the workshop focused on human–robot interaction and deployment in social environments. Michael Gienger (Honda Research Institute Europe) discussed the use of large language models to drive robot action and interaction, sharing early experiments in LLM-based planning and highlighting challenges in grounding language in real-world situations. He emphasized the importance of keeping humans at the center, pointing toward future multimodal policies that can detect failures and support explainable interaction.

Xavier Alameda-Pineda

Xavier Alameda-Pineda (Inria) addressed similar challenges from a multimodal perspective. He presented work on diffusion-based, unsupervised audio–visual speech enhancement, followed by research on text-conditioned safe social navigation. Drawing on a completed EU project evaluating a conversational robot in a gerontological day-care hospital, he offered rare insights into the practical difficulties of large-scale, in-situ evaluation, where technical performance and human factors are deeply intertwined.

Where AI for robotics and Spatial AI are heading

Across domains and approaches, several shared conclusions emerged:

Spatial AI is becoming more dynamic: time, motion and interaction are no longer secondary concerns but core components of representation.

Scale cuts both ways: large models and datasets drive progress, but efficiency, hardware and system design increasingly constrain what is deployable.

Learning and structure must coexist: foundation models thrive when paired with geometry, control and model-based reasoning.

Real-world evaluation matters: deployment in cities, cars, and hospitals exposes challenges that benchmarks alone cannot capture.

Taken together, the workshop painted the field of Spatial AI as moving beyond isolated advances toward integrated systems that can reason, act and adapt in the real world over time, at scale and in the presence of humans.