We focus on parsing procedural text into fine-grained flow graphs that encode actions and entities, as well as their interactions. Unlike prior two-stage approaches, we propose an end-to-end multi-task framework that simultaneously performs tagging and graph generation, and use noisy silver annotations from unlabeled recipes to improve learning with limited annotated data.

We introduce flow graph to video grounding for weakly supervised multi-step localization. Instead of relying on human-provided step order for every video, we use procedural flow graphs to represent valid step orderings and infer the ordering and localization jointly.

We explore how Abstract Meaning Representation can address limitations of scene-graph-like visual abstractions. The work shows that existing text-to-AMR parsers can be repurposed toward image-based AMR representations and points toward richer visual scene understanding.

f-Cal improves uncertainty calibration by enforcing constraints across batches rather than only per sample. The method produces better calibrated uncertainty estimates for robot perception tasks such as object detection and monocular depth estimation.

We survey the landscape of probabilistic object detection, identify key shortcomings, and outline promising directions toward accurate, reliable, and consistent uncertainty estimates in autonomous driving perception systems.

MapLite pilots a vehicle to a destination using sparse public topometric maps, onboard road segmentation, and local planning. The system navigates intersections without GPS-like external localization or detailed prior maps.

We propose an obstacle-avoidance and object-following method that accounts for uncertainty in robot and obstacle states by optimizing control with a bounded-risk measure based on overlap between Gaussian distributions.

We present an end-to-end differentiable active localization system that learns both perception and planning in simulation, then transfers to real robot hardware with no refinement.

We pose intersection detection as binary classification over a sequence of frames and use a recurrent convolutional model to capture temporal consistency in monocular road images.

We combine pixel-wise semantic segmentation with depth and color information from stereo cameras, using dense CRFs to improve curb detection under varied road conditions.