arXiv subjects

Mark Carlson

Publications and source records attributed to Mark Carlson.

4 recordsLinked to original sources

arXiv2026

Cosmos 3: Omnimodal World Models for Physical AI

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. By supporting highly flexible input-output configurations, Cosmos 3 seamlessly unifies critical modalities for Physical AI -- effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. Our evaluation demonstrates that Cosmos 3 establishes a new state-of-the-art across a diverse suite of understanding and generation tasks, demonstrating omnimodal world models as scalable, general-purpose backbones for embodied agents. Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written. To accelerate open research and deployment in Physical AI, we make our code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under the Linux Foundation's OpenMDW-1.1 License at https://github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3. The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos3.

cs.CV↗

arXiv2025

GraspGen: A Diffusion-based Framework for 6-DOF Grasping with On-Generator Training

Grasping is a fundamental robot skill, yet despite significant research advancements, learning-based 6-DOF grasping approaches are still not turnkey and struggle to generalize across different embodiments and in-the-wild settings. We build upon the recent success on modeling the object-centric grasp generation process as an iterative diffusion process. Our proposed framework, GraspGen, consists of a DiffusionTransformer architecture that enhances grasp generation, paired with an efficient discriminator to score and filter sampled grasps. We introduce a novel and performant on-generator training recipe for the discriminator. To scale GraspGen to both objects and grippers, we release a new simulated dataset consisting of over 53 million grasps. We demonstrate that GraspGen outperforms prior methods in simulations with singulated objects across different grippers, achieves state-of-the-art performance on the FetchBench grasping benchmark, and performs well on a real robot with noisy visual observations.

cs.RO↗

arXiv2025

CHIME/FRB Outriggers: Design Overview

The Canadian Hydrogen Intensity Mapping Experiment (CHIME) has emerged as the world's premier facility for studying fast radio bursts (FRBs) through its fast transient search backend CHIME/FRB\@. The CHIME/FRB Outriggers project will augment this high detection rate of 2--3 FRBs per day with the ability to precisely localize them using very long baseline interferometry (VLBI). Using three strategically located stations in North America and deploying recently developed synoptic VLBI observing techniques, the Outriggers will provide $\sim 50$~milliarcsecond localization precision for the majority of detected FRBs. This paper presents an overview of the design and implementation of the Outriggers, covering their geographic distribution, structural design, and observational capabilities. We detail the scientific objectives driving the project, including the characterization of FRB populations, host galaxy demographics, and the use of FRBs as cosmological probes. We also discuss the calibration strategies available to mitigate ionospheric and instrumental effects, ensuring high-precision localization. With two stations currently in science operations, and the third in commissioning, the CHIME/FRB Outriggers project is poised to become a cornerstone of the FRB field, offering unprecedented insights into this enigmatic cosmic phenomenon.

astro-ph.HE↗

arXiv2024

CHIME/FRB Outriggers: KKO Station System and Commissioning Results

Localizing fast radio bursts (FRBs) to their host galaxies is an essential step to better understanding their origins and using them as cosmic probes. The CHIME/FRB Outrigger program aims to add VLBI-localization capabilities to CHIME, such that FRBs may be localized to tens of milliarcsecond precision at the time of their discovery, more than sufficient for host galaxy identification. The first-built outrigger telescope is KKO, located 66 kilometers west of CHIME. Cross-correlating KKO with CHIME can achieve arcsecond-scale localization in right ascension while avoiding the worst effects of the ionosphere. This paper presents measurements of KKO's performance throughout its commissioning phase, as well as a summary of its design and function. We demonstrate KKO's capabilities as a standalone instrument by producing full-sky images, mapping the angular and frequency structure of the primary beam, and measuring feed positions. To demonstrate the localization capabilities of the CHIME -- KKO baseline, we collected five separate observations each for a set of twenty bright pulsars, and aimed to measure their positions to within 5~arcseconds. All of these pulses were successfully localized to within this specification. The next two outriggers are expected to be commissioned in 2024, and will enable subarcsecond localizations for approximately hundreds of FRBs each year.

astro-ph.IM↗