Physical AI Takes Shape: Training Robots from Video

Friday afternoon, and the robotics world is moving fast. There's real money flowing into physical AI now-not the speculative kind, but funding tied to actual deployments that work in messy, real-world environments. That shift matters.

Robots Learning from Video

Rhoda AI emerged from stealth this week with $450 million in Series A funding, and their approach is worth paying attention to. Instead of teaching robots through manual teleoperation or pre-programmed trajectories, they're training systems on internet-scale video. Their Direct Video-Action model watches how the world moves, learns physics and interaction patterns, then operates in closed loop-continuously predicting what comes next, taking action, observing the result, and adjusting. It's elegant. The company claims robots can learn new tasks from as little as ten hours of teleoperation data, which would be transformative if it holds at scale. They've already demonstrated autonomous operation in production manufacturing, completing component-processing workflows without human intervention.

Alongside Rhoda, MassRobotics announced the second cohort of their Physical AI Fellowship, backed by AWS and NVIDIA. Nine startups are now in the program, working on everything from agricultural automation to humanoid robotics. The names read like a map of where the industry is heading: Burro (autonomous farm robots), Config (multi-arm learning), Telexistence (retail automation in Japan), and others tackling real industrial problems. What stands out is the explicit focus on moving from prototypes to enterprise deployments. These aren't research projects anymore.

Building the Infrastructure for Agentic Systems

Zooming out, though, there's a parallel story unfolding in how we handle data at scale. Simon Hørup Eskildsen of Turbopuffer spent nearly a decade at Shopify scaling databases, and he noticed something: when AI agents start doing real work, they don't make one search query and wait for an answer. They fire off dozens of parallel queries, looking for context from different angles. That changes everything about how you build search infrastructure. It's no longer about single, carefully-chosen retrievals. It's about handling massive concurrent load from systems that can think in parallel. He's already cut query pricing by 5x to accommodate this shift, and expects to cut it further as agentic workloads become standard.

The deeper insight: physical AI and intelligent agents both need the same infrastructure-fast, cheap retrieval systems that can handle variable, unpredictable access patterns. Whether it's a robot vision system querying learned representations or an agent searching across code or documents, the workload is evolving away from the patterns databases were built to optimize for.

These aren't disconnected stories. They're early signals of a reconfiguration happening in parallel across hardware (robots), software (agents), and infrastructure (search and databases). The money is flowing toward teams that can handle real-world complexity without retreating to controlled environments. That's the shape of things Going ahead.