Cube: A Roblox View of 3D Intelligence

Foundation models trained on vast amounts of data have demonstrated remarkable reasoning and generation capabilities in the domains of text, images, audio and video. Our goal at Roblox is to build such a foundation model for 3D intelligence, a model that can support developers in producing all aspects of a Roblox experience, from generating 3D objects and scenes to rigging characters for animation to producing programmatic scripts describing object behaviors. We discuss three key design requirements for such a 3D foundation model and then present our first step towards building such a model. We expect that 3D geometric shapes will be a core data type and describe our solution for 3D shape tokenizer. We show how our tokenization scheme can be used in applications for text-to-shape generation, shape-to-text generation and text-to-scene generation. We demonstrate how these applications can collaborate with existing large language models (LLMs) to perform scene analysis and reasoning. We conclude with a discussion outlining our path to building a fully unified foundation model for 3D intelligence.

@article{roblox2025_2503.15475,
  title={ Cube: A Roblox View of 3D Intelligence },
  author={ Foundation AI Team Roblox and Kiran Bhat and Nishchaie Khanna and Karun Channa and Tinghui Zhou and Yiheng Zhu and Xiaoxia Sun and Charles Shang and Anirudh Sudarshan and Maurice Chu and Daiqing Li and Kangle Deng and Jean-Philippe Fauconnier and Tijmen Verhulsdonck and Maneesh Agrawala and Kayvon Fatahalian and Alexander Weiss and Christian Reiser and Ravi Kiran Chirravuri and Ravali Kandur and Alejandro Pelaez and Akash Garg and Michael Palleschi and Jessica Wang and Skylar Litz and Leon Liu and Anying Li and David Harmon and Derek Liu and Liangjun Feng and Denis Goupil and Lukas Kuczynski and Jihyun Yoon and Naveen Marri and Peiye Zhuang and Yinan Zhang and Brian Yin and Haomiao Jiang and Marcel van Workum and Thomas Lane and Bryce Erickson and Salil Pathare and Kyle Price and Steve Han and Yiqing Wang and Anupam Singh and David Baszucki },
  journal={arXiv preprint arXiv:2503.15475},
  year={ 2025 }
}