A startup spun out of Israeli creative technology firm Lightricks is deploying synthetic video data to train robotic systems, entering one of the most competitive and capital-intensive frontiers in artificial intelligence. The company, operating under the name Physical Intelligence adjacency, aims to address a fundamental bottleneck in robotics development: the scarcity of high-quality real-world training data. As global investment in Israel’s tech sector continues to attract international attention, the spinoff represents a notable pivot from Lightricks’ core business in AI-powered visual content tools.
According to a report by Globes, the Israeli financial news outlet, the venture is built on the premise that video — both real and synthetically generated — can serve as a scalable substitute for the expensive and time-consuming process of training robots through physical trial and error in the real world. The approach mirrors techniques used in large language model development, where vast datasets substitute for direct human instruction, and applies that logic to the physical world of motion, manipulation, and spatial reasoning.
The Data Scarcity Problem Driving Demand
Training robots to perform even rudimentary tasks — grasping irregular objects, navigating unpredictable surfaces, responding to human movement — requires enormous volumes of demonstration data. Unlike natural language, where billions of text documents exist online, physical interaction data is sparse, proprietary, and prohibitively costly to generate at scale. Industry estimates suggest that leading robotics firms spend tens of millions of dollars annually on data collection infrastructure alone, with some programs requiring thousands of hours of supervised human demonstration per skill set.
The Lightricks spinoff’s proposition is that generative video models, trained on existing footage of human and mechanical movement, can produce synthetic training datasets that approximate real-world conditions closely enough to be useful. If the approach proves viable, it could dramatically compress the time and capital required to bring capable robotic systems to deployment. The technology builds directly on Lightricks’ existing expertise in video generation and AI-driven visual synthesis, giving the new venture a meaningful head start over competitors building such capabilities from scratch.
Strategic Position in a Rapidly Consolidating Market
The robotics training data market is attracting growing corporate and venture interest, with major technology groups including Alphabet, Amazon, and a cohort of well-funded startups all racing to establish defensible positions. Physical AI — the discipline of teaching machines to operate reliably in unstructured real-world environments — has been widely identified as the next major commercial frontier after large language models, with some analysts projecting the addressable market for robotic systems to exceed several hundred billion dollars within the decade.
For Lightricks, the spinoff strategy allows the parent company to monetise its video AI infrastructure in an adjacent vertical without diverting its core product roadmap, which remains focused on tools for creative professionals and content producers. The structure also makes the new entity more attractive to sector-specific investors, who may be reluctant to back a creative software company diversifying into deep robotics research. Lightricks has previously raised significant capital, including a funding round that valued the company at approximately 1.8 billion dollars, positioning it as one of Israel’s most prominent technology unicorns.
The spinoff’s success will ultimately hinge on whether synthetically generated video data can meet the fidelity thresholds that robotic learning systems require — a question the broader research community has not yet resolved. Early results from comparable efforts at institutions including Stanford and Carnegie Mellon have been promising but limited to controlled environments. Commercialising that capability at scale, with consistent performance across real-world variables, remains an open engineering challenge that the new company will need to demonstrate before attracting the institutional partnerships and enterprise contracts necessary to sustain growth.