CAIRO and ARPG have published a paper at the IEEE International Conference on Robotics and Automation (ICRA 2025):

Online Diffusion-Based 3D Occupancy Prediction at the Frontier with Probabilistic Map Reconciliation
by Alec Reed, Lorin Achey, Brendan Crowe, Bradley Hayes, and Christoffer Heckman.

In this work, the team introduces real-time diffusion-based 3D occupancy prediction methods that allow robots to reason about unobserved regions of their environment—critical for autonomous navigation and exploration. By modifying the SceneSense framework, the approach reduces runtime by 73%, enabling predictions not just around the robot but across the entire map.

Key contributions include:

• A probabilistic map update rule for merging predicted occupancy into running maps, preserving coherence and accuracy.
• The ability to generate occupancy predictions at map frontiers, improving prediction quality by 71% compared to prior methods.
• Integration of this method on a real quadruped robot (Boston Dynamics Spot), demonstrating enhanced frontier exploration and decision-making in unmapped environments.

Figures in the paper illustrate how the system can fill in missing geometry at hallway intersections, stairwells, and occluded regions, supporting safer and more efficient exploration.

📄 Full paper PDF: CAIRO Lab Link 🔗 Project page & ROS node: https://arpg.github.io/scenesense/

This work was supported by the National Science Foundation (Award #2339328) and highlights the potential of generative AI techniques in robotics to improve map inference and exploration autonomy.