To generate video with artificial intelligence, a user typically submits a prompt, waits, and then receives a finished clip that cannot be altered. Visko Platform Inc. today opened public access to Orbis, a foundation model that streams worlds in real time with persistent memory, interactivity, and physics-grounded generation for unlimited duration. The company has also raised $10 million in pre-seed funding led by Llama Ventures.
“Every real-time system before this faced the same wall: The longer a world runs, the more it falls apart,” stated Qing (Will) Yin, founder and CEO of Visko. “We built Orbis to hold together — physically, visually, and narratively — for as long as you want to stay inside it. Starting today, anyone can see that for themselves at visko.ai.”
“Most video models optimize every frame of a clip at once and show nothing until the whole thing resolves,” Yin told The Robot Report. “Orbis finishes the first frame, streams it, then generates the next conditioned on what came before, closer to next-token prediction in a large language model.”
Visko was founded in 2025 by Yin, a Stanford Ph.D. in computational mathematics and mechanics who spent three years as a researcher at Apple. The Sunnyvale, Calif.-based startup‘s 16-person team draws from Apple, Google DeepMind, Meta, Amazon, and Tesla, and it is working with Google.
Editor’s note: Physical AI and enabling technologies will be among the session track topics discussed at RoboBusiness 2026, which will be on Oct. 20 and 21 in Santa Clara, Calif. Register now to attend.
Register now and save on your pass to RoboBusiness 2026Orbis runs world models continuously
Visko said its Live Models are a new class of world models that run rather than render. Where conventional AI video generates a clip and stops, the company said its Orbis foundation model generates video continuously and streams it live.
Users can change the prompt at any moment during generation and see the update reflected instantly while the video keeps streaming. “The next paradigm of AI is not a model you call, but a process that runs in step with the world itself,” Visko claimed.
According to the company’s technical report, Orbis delivers 4K video at 24 frames per second in real time and sustains hour-scale generation without quality or color drift. “Generation is grounded in physical dynamics: A latent world model scores candidate futures during inference, steering output toward physically plausible motion and object dynamics,” said Visko.
“Change the prompt mid-run, and the world updates around it instead of restarting,” said Yin. “That makes it usable for simulation and for training footage that is unsafe or expensive to stage. In short, we are the world’s first real-time 4K steerable video generation model, enabling use cases in interactive robotic training.”
“In addition to generating videos up to an hour long and enabling real-time commands, Orbis can generate videos in real time,” said Yin. “We designed the memory structure to directly embed physics frame by frame. This actually requires less compute, and egocentric data requires long-form video without any breaks.”
In the report’s evaluations, Orbis leads compared real-time and long-video systems on DOVER aesthetic and technical quality and on VideoAlign visual and motion quality. Visko said it recorded the highest overall-preference and temporal-stability ratings in a long-form human preference study across eight systems.
Orbis is a real-time, interactive long video-generation system. Source: Visko
Visko has more work to do
The report also disclosed dimensions where Orbis trails other systems, with full methodology and results published. There is more work to be done, acknowledged Yin.
“On the research front, we need better design of the memory structure,” he said. “For example, the video can maintain even better consistency across the generation.”
“Secondly, we would like to add more detailed action description into the data collecting, like for training those video models,” Yin added. “Right now, I can ask a robotic arm to do something, but I don’t have a very detailed description. If I give a very detailed instruction, like raise the arm 45 degrees, we would like to incorporate more accurate action description data into training this model.”
“On the engineering side, we have to scale up the model in terms of the model and the data size,” he said. “We’re focused on audio and video data at the moment. We’re still exploring injecting other sensor data, like tactile. For a lot of deformable objects, we can interpret the forces from the video, but we’re looking for partners focusing on robotic sensors and applications. We’re talking to some hardware companies.”
Top: Quality, alignment, and motion measurements on one- to three-minute event-based benchmark. Bottom: Overall long-form video Arena point estimates from the human preference study. Orbis 1.0 attained the highest overall score; the table provides the dimension-level comparison and citations for all baselines. Source: Visko
Advisor cites potential applications
Advisors include Michael I. Jordan of the University of California, Berkeley; Steve WaiChing Sun of Columbia University; and Mengye Ren of New York University. Distribution is handled through Visko’s partner Reactor.
“Visko is pursuing one of the most ambitious directions in video AI: live, stateful models that continuously perceive, predict, and interact with the world in real time while maintaining physical consistency,” said Jordan, the chairman of Visko’s advisory board.
“Visko delivers a state-of-the-art architecture that has a meaningful technological lead beyond conventional short-clip video generation,” he added. “The same core capability could support a remarkably broad range of commercial opportunities, from robotics and physical simulation to gaming, live commerce, education, and real-time creative media.”
Orbis could also help provide environments for testing corner cases for safer autonomous vehicles and humanoid robots, said Yin.
“Some people people estimate that general-purpose robots will be in homes in five to 10 years, but from my communications with people in the field, that could take some time, just like autonomous driving,” he noted. “My goal is to find commercialization opportunities on the way.”
The data for testing Orbis 1.0 included raw video collection of real-world and generated domains. Source: Visko
Tell Us What You Think!