Research Scientist at Luma AI
Multimodal AI · Embodied AI · Generative Models · Recursive Self-Improvement
I am a Research Scientist at Luma AI, where I work on multimodal and embodied AI. I am a founding researcher on Project Opal, Luma's open physical AI initiative, and a core contributor to Uni, Luma's unified multimodal reasoning and generation model. I received my Ph.D. in Computer Science from the University of Maryland, College Park, advised by Prof. Abhinav Shrivastava, in 2025.
My research focuses on model training and data curation for multimodal and embodied AI, visual generation and video representation learning, and recursive self-improvement for embodied agents. Most recently, I led the real-robot work for SEAR, including harness design and experiments.
A benchmark and framework for studying how LLM agents improve robot policies through execution feedback and carry skills and lessons across tasks, under controlled time budgets. I led the real-robot work for SEAR, including harness design and experiments.
Uni-1 is a multimodal reasoning model that understands intention, responds to direction, and generates pixels.
Synthesizes videos by diffusing implicit neural representation weights, then decoding frames from their indices.
Adds image generation to pre-trained multimodal understanding models with a Mixture-of-Transformers architecture and semantic-to-pixel visual tokens.
Learns holistic video tokens with an autoregressive prior to align tokenization with downstream autoregressive video generation.
