Covariant was founded by Peter Chen, Pieter Abbeel, Rocky Duan, and Tianhao Zhang, all affiliated with UC Berkeley’s AI research program. The company builds AI for robotic manipulation, enabling robots to pick, place, and handle virtually any object in warehouse and logistics environments using learned rather than programmed behaviors.
Core Products
RFM-1 (Robotics Foundation Model) is Covariant’s multimodal model that can reason about the physical world, understand natural language instructions, and predict robot actions. The Covariant Brain powers robotic picking systems that handle millions of unique objects without explicit programming for each item. The platform deploys across diverse logistics environments including warehouses, distribution centers, and fulfillment operations.
Competitive Position
Covariant’s approach of building foundation models for robotics, analogous to what GPT is for language, positions it at the frontier of embodied AI. The company has real-world deployments processing millions of picks, providing training data that improves its models. Competition from Physical Intelligence, Amazon Robotics, and others is intensifying, but Covariant’s deployed experience and academic pedigree provide advantages.
Recent Developments
Covariant unveiled RFM-1, which can understand and generate both language and physical actions. The company has expanded deployments across major logistics providers and e-commerce companies. Reports in 2024 suggested Amazon was in discussions to acquire or deeply partner with Covariant, reflecting the strategic value of physical AI capabilities.
Outlook
Covariant sits at the intersection of two massive trends: AI foundation models and logistics automation. The global warehouse robotics market is expected to reach tens of billions in value. Covariant’s challenge is scaling deployments while maintaining technical leadership against both well-funded startups (Physical Intelligence raised $400M) and tech giants with robotics ambitions.