An object-relative control framework for visual navigation using a relative 3D scene graph, enabling more flexible, embodiment-invariant path planning and control that generalizes better across tasks and real-world environments than traditional image-relative methods.

A global localization system that uses VLMs to build sparse, semantic maps. It achieves localization accuracy comparable to dense LiDAR methods while using 500x fewer points. Deployed in production by Ati Motors.

Proposes a new framework for evaluating LLM-based forecasters by measuring their logical consistency, introducing an "arbitrage metric" that correlates strongly with future forecasting accuracy, even when the ground truth is unknown.

A framework to build an object-based global map, and enables localization over it through a novel dual path transformer architecture that performs multi-modal instance re-identification.