Brandon B. May

TL;DR: A training-free, test-time framework that reconstructs simulation-ready assets from a single RGB-D image using 3D generative and vision-language models, then launches thousands of parallel physics rollouts with evolutionary search to optimize object arrangements for language-specified tasks. Achieves 67% success on real-robot hardware, surpassing baselines using off-the-shelf foundation models.

TL;DR: We invert the traditional sim-to-real paradigm by building a dynamic digital twin powered by an Embodied Gaussian simulator that synchronizes with the real world at 60Hz. Instead of training on real hardware or adapting simulations post-hoc, policies always execute on a virtual robot while the physical robot mirrors its joint states. Continuous real-world measurements correct the simulation on the fly, and we validate the approach on long-horizon manipulation tasks like PushT, showing tight alignment between virtual evaluations and physical results.

TL;DR: PIEGraph learns physically grounded dynamics for rigid and deformable objects from just a few minutes of human interaction data. It combines a physics-informed spring-mass prior with an action-conditioned equivariant graph neural network, maintaining physical plausibility (no interpenetration, shape preservation) where standard particle-based models break down. Demonstrated on ropes, cloth, stuffed animals, and rigid bodies for robotic planning.

TL;DR: Theia distills multiple off-the-shelf vision foundation models (DINOv2, CLIP, SAM, Depth-Anything, and more) into a single compact model optimized for robot learning. The result outperforms any individual teacher model while being smaller and requiring less training data. We also find that higher entropy in feature norm distributions correlates with better downstream robot performance, offering a practical proxy for representation quality. Pre-trained models available on Hugging Face.

TL;DR: Sancdifi defends against backdoor attacks by using a denoising diffusion model to degrade and recover images, with saliency-based conditioning that concentrates the diffusion on the most important regions. This strips backdoor triggers while preserving legitimate features. Crucially, it works as a black-box defense with no access to the target model's internals.

TL;DR: We release a first-of-its-kind dataset of real, fully synthetic, and partially manipulated overhead imagery for forensic research. The synthetic images are generated from a custom diffusion model trained across multiple zoom levels and data sources, enabling research into detecting and localizing manipulated satellite imagery.

TL;DR: We introduce the first comprehensive benchmark for explainable face recognition, including “the inpainting game,” a standardized evaluation protocol of 3,648 triplets where facial features are synthetically modified to create ground truth. We propose two new attention methods, subtree EBP and DISE (Density-based Input Sampling for Explanation), which significantly outperform prior techniques at revealing which facial regions drive a network's matching decisions.