Spirit AI has raised $280 million to scale the deployment of its general-purpose embodied models, as the industry shifts toward Scaling Law-driven VLA architectures. The funding comes from a diverse group of global financial and strategic investors and supports the company’s push to expand large-scale physical AI systems trained on real-world data.
Based in Beijing, Spirit AI is developing what it describes as a universal robotic brain by scaling models with diverse human video and wearable sensor data. The company’s approach aligns with global peers such as Google DeepMind and Physical Intelligence in leveraging massive datasets for physical reasoning. Its core team, drawn from UC Berkeley, Tsinghua University, and Peking University, averages under age 30 and combines multimodal large language model research with robot learning and industrial deployment.
Central to Spirit AI’s strategy is what it calls “dirty data” — diverse, unstructured, and non-pre-scripted interaction data gathered from real-world environments. While some competitors focus on highly curated datasets, Spirit AI argues that over-curation can limit model generalization and lead to performance plateaus. Instead, the company prioritizes real-world complexity as a path to building models with broader common sense capabilities.
Spirit AI has accumulated more than 200,000 hours of interaction data and aims to surpass 1 million hours by the end of 2026. Through proprietary wearable data collection devices, the company says it has reduced data acquisition costs by 90% compared to traditional teleoperation methods. In January 2026, its Spirit v1.5 model topped the RoboChallenge global leaderboard, demonstrating state-of-the-art generalization performance comparable to leading embodied AI systems.
The company has also deployed its VLA models in industrial settings, including on production lines at CATL, the world’s largest battery manufacturer. In that environment, Spirit AI-powered robotic agents manage flexible wire harnesses, a task historically challenged by material unpredictability. According to the company, the system achieves a 99% or higher success rate while matching the precision and cycle times of skilled human workers in complex manufacturing processes.
Spirit AI says its broader goal is to build a “Universal Brain” for next-generation robotics by bridging simulation and real-world deployment. By integrating general-purpose embodied models into practical workflows, the company aims to accelerate the adoption of versatile robotic agents across modern industrial environments.
KEY QUOTES
“Dirty data is the key to scaling VLA models.”
Yang Gao, Co-founder & Chief Scientist, Spirit AI
