1.36B 'world' model trained on robot videos
Vizuara unveiled a 1.36B-parameter World Foundation Model trained on robot video data aimed at action imagination and simulation—positioning video‑trained foundation models as a shortcut to embodied capabilities. The release highlights a trend of trading larger simulation datasets for cheaper hardware‑dependent collection. (x.com)
Vizuara describes itself as an AI education and research startup founded in 2023 and lists founders Sreedath Panat, Raj Abhijit Dandekar and Rajat Dandekar; Tracxn’s company profile notes the firm is based in Bengaluru and remains unfunded. (tracxn.com) The organization has been running a “Modern Robot Learning” bootcamp that began March 21, 2026 and advertises hands‑on modules that include building video tokenizers, captioning pipelines, diffusion policies and world‑model training from scratch. (robotlearningmastery.vizuara.ai) Vizuara’s public posts and course materials explicitly discuss pretraining on large web/video corpora and fine‑tuning on small amounts of robot data as a pathway to transfer physical skills across embodiments. (youtube.com) A public Vizuara GitHub account and a modest Hugging Face presence exist, but neither the organization’s GitHub nor its Hugging Face pages list a publicly archived world‑model weights release or an open model repository for the new world model. (github.com) The release aligns with parallel industry work: NVIDIA and other labs have promoted “world foundation models” and sim‑and‑real co‑training for embodied AI, while academic projects (PhysWorld, 1X and others) have demonstrated robot relevant pretraining from human and internet video. (developer.nvidia.com) Vizuara’s messaging positions video‑first world models alongside VLA and diffusion‑policy toolchains — an approach that other groups have argued can reduce the need for expensive, hardware‑heavy robot data collection by leveraging large video corpora and simulators in co‑training. (youtube.com)
