Episode Summary
Executive Summary: Peter Chen explains why robotics needs foundation models built on real-world data, not just simulation. Covariant focuses on warehouse manipulation because it offers immediate commercial value, abundant task diversity, and a path to collect the high-quality data needed for more general robot intelligence. The conversation contrasts today’s rigid industrial robots with a future of adaptive, reliable, AI-powered systems.
Main Topics: Why robotics attracted Peter Chen (Priority: 5/5): Chen links robotics to two AI interests from research: unsupervised learning/foundation models and reinforcement learning/decision-making. Robotics combines both because robots must perceive the world and act in it with consequences. Covariant’s company thesis: data first, products second (Priority: 5/5): Covariant was founded on the belief that robotics foundation models require large amounts of real robot data, which can only be collected by deploying useful systems in production. The business exists to generate the data needed to improve the models. Why warehouse robotics is the initial beachhead (Priority: 5/5): Warehouses are full of repetitive yet highly variable manipulation problems, and labor shortages plus high turnover make automation economically compelling. Covariant starts here because the task diversity is large and the ROI is clear. What makes robots 'smart' vs. 'dumb' today (Priority: 5/5): Chen argues that most deployed robots are still pre-programmed and rigid. The opportunity is not incremental automation of existing uses, but expanding into new use cases requiring adaptation, perception, and intelligent action. Grounding, precision, and action data (Priority: 5/5): He distinguishes internet-scale multimodal grounding from robotics data needs: robots require sub-millimeter precision and action-outcome feedback that captioned images/videos rarely provide. That missing physical interaction data is central to progress. Simulation vs. real-world learning (Priority: 4/5): Covariant uses simulation as augmentation, not replacement. Chen says contact-rich manipulation, deformable objects, and the effort required to model huge item inventories make simulation insufficient as the primary training source. Future applications, humanoids, and safety (Priority: 4/5): Chen expects industrial robotics to mature before consumer/home robots because ROI is stronger and safety is more manageable. Humanoids are attractive long-term, but near-term deployment will remain in industrial settings with strong safety constraints.
Key Arguments: Robotics is the intersection of unsupervised learning and reinforcement learning: robots must both understand the world and take actions in it. Covariant was not built after robotics became commercially ready; it was built because Chen believed the only way to create robotics foundation models was to deploy useful robots and collect production data. The future of robotics will be shaped by whoever has the most real robotics data, not by simulation alone. Industrial warehouses are a strong starting point because they have high labor turnover, enormous task diversity, and immediate economic value from automation. Most current robots are 'dumb' in the sense that they are rigidly pre-programmed; the real opportunity is AI-driven adaptability in diverse environments. Internet grounding is useful but insufficient for robotics because it lacks precision and action-outcome data at the level needed for physical manipulation. Simulation cannot fully replace real-world data in manipulation because contact dynamics are hard to model and warehouse item diversity is enormous. The next major leap in robotics will require both generality and very high reliability, since even a small failure rate can be catastrophic in the physical world. Humanoid robots are appealing because human environments are designed for human bodies, but industrial robots will likely be the first major commercial success. Safety is more tractable in industrial settings because existing certification, cages, and controller rules constrain what the robot can do.
Data Points: Company size: About 200 people - Covariant’s current scale five years into the company. Geographic customer mix: Roughly half in Europe, half in North America - Distribution of Covariant’s customers. Deployment footprint: Across 3 continents and more than 10 countries - Where Covariant robots are currently deployed. Warehouse turnover rate: More than 100% year-over-year turnover - Used to illustrate labor shortage and operational pain in warehouse work. Robot prevalence: 99%+ of deployed robots are dumb robots - Chen’s estimate that most robots in the world are pre-programmed and not intelligent. Task coverage: Hundreds of thousands of unique items - Example of the diversity a warehouse pick-and-pack robot must handle. Human oversight ratio: One person overseeing 10, 20, or 30 robots - Chen’s vision for robotics-augmented warehouses. Industrial usage intensity: 24/7 - Why industrial robots justify hardware investment better than home robots. Home robot usage: About 2 hours a week - Why consumer robotics is harder to justify economically in the near term.
Pivotal Quotes: "the future of AI is going to be the future of foundation models" — Peter Chen: Explaining the original thesis behind Covariant and why robotics should follow the same scaling pattern as language and multimodal AI. "the future of robotics would be built by whoever that has most robotics data" — Peter Chen: Defining Covariant’s core strategic bet and why the company prioritizes real-world deployment. "you really need a large amount of high-quality data to densely cover this robotic field" — Peter Chen: Describing why reliability and breadth of real-world experience matter for the 'ChatGPT moment' in robotics.
Implications: Robotics progress will likely come from data-rich industrial deployments, not simulation-only research. Expect smarter warehouses first, humanoids later, and consumer robots last, with safety and reliability becoming central differentiators.