NU · World Model & Embodied AI

We build and benchmark world models, embodied agents, and world-action models — systems that simulate, predict, and act in the physical world.

Our public releases include video-generation benchmarks, physics-aware world models, robot policies, action models, datasets, and open evaluators grounded in human judgments of the physical world.

Part of the Moxin Open AI Ecosystem

NU World Model & Embodied AI is part of the Moxin Organization open AI ecosystem, extending Moxin’s work from language models, agents, applications, and efficient inference into world models and physical intelligence.

Moxin Website · Moxin GitHub · Moxin Hugging Face

Current Releases

PhyGround — the ruler

PhyGround is a benchmark for evaluating the physical plausibility of text-and-image-to-video generation.

It includes 250 curated prompts, a taxonomy spanning 13 physical laws across solid-body, fluid, and optical domains, and a quality-controlled human study with 459 annotators and more than 37,000 fine-grained labels.

We also release PhyJudge-9B, an open vision-language evaluator trained on the human ratings.

Project Page · GitHub · Dataset · PhyJudge-9B

PhyWorld — a model trained against the ruler

PhyWorld is a physics-aware video-generation world model post-trained from Wan2.2-I2V-A14B.

Its two-stage training combines flow-matching fine-tuning for temporal coherence with preference optimization using physics judgments derived from the PhyGround human-annotation pool.

Project Page · GitHub · Model

Flash-WAM — efficient world-action models

Flash-WAM is a modality-aware distillation framework that accelerates joint video-action world models while preserving embodied-task performance.

GitHub · Browse Releases

A Shared Research Loop

PhyGround and PhyWorld form a shared research loop: the human annotations used to evaluate generated videos also provide the preference signals used to train a more physically coherent world model.

More releases in robot policies, world-action models, efficient embodied intelligence, and physical-reasoning benchmarks are on the way.