General Robotics is deliberately walking away from the single-brain narrative dominating physical AI, CTO and co-founder Sai Vemprala said on The Robot Report Podcast Episode 263 published September 25. Instead of training one monolithic foundation model, the Redmond, Washington-based startup is composing specialized skills into a modular runtime that any robot body can subscribe to.
The GRID Bet
General Robotics' flagship product, GRID, is what Vemprala calls a "General Robot Intelligence Development" platform. It ships 40-plus pre-trained skills — grasping, navigation, perception, dexterous manipulation — that developers can chain together and deploy in as little as 15 minutes across cloud, on-prem or edge. GRID plugs into NVIDIA Isaac Sim for training and can inference more than 25,000 concurrent requests, aimed at fleet-scale operators.
Why Modular
Vemprala, a former Microsoft Research scientist with a Texas A&M robotics PhD, argues that swapping in a purpose-built skill is faster, cheaper and safer than retraining a foundation model whenever a robot needs a new capability. "Every skill has its own guarantees, its own compute footprint and its own upgrade cadence," he said. That is a pointed contrast with Skild AI's S1 self-play brain and Nvidia's own end-to-end foundation model push.
Cross-Form-Factor By Design
Because skills are decoupled from the body, GRID targets humanoids, arms, quadrupeds, wheeled AMRs and even aerial platforms with the same runtime. That is the same mixed-fleet future Agility Robotics is chasing with a wheeled Digit companion and Toyota is engineering into its 400,000-robot factory rollout. General Robotics' pitch to OEMs and end-users is that they will not need to pick a stack per robot body.
Weekly News Also Discussed
The same episode covered Qualcomm's PickNik acquisition, Boston Dynamics' latest robot demos, and Cognex's most recent M&A activity — a sign that vendors are consolidating around competing definitions of the "robot operating layer" that GRID is trying to become.
Reporting based on coverage from The Robot Report Podcast.