General Robotics: AI agents for vastly faster robot training

In this talk

1. 1:01

   The gap between research and deployment

   Kapoor says robotics systems have not kept pace with advances in deep learning, perception models, and cloud tools. Unlike software machine learning, which gained shared platforms such as PyTorch and TensorFlow, robotics still takes specialized engineering to bring research into field use. He describes AutoEng as a team of virtual robotics specialists intended to help close that gap.

2. 5:07

   GRID connects a fragmented robotics stack

   Kapoor explains that robotics work spans hardware, simulation, data collection, model tuning, testing, and deployment, with tools and vendors often isolated from one another. General Robotics built GRID as a robotics monorepo with robot models, simulators, AI models, deployment processes, and lower-level control code, plus interfaces that agents can use. He says agents could combine simulation components for a specific task, potentially avoiding months of custom work.

3. 14:43

   Supporting fleets of different robots

   Kapoor expects large enterprises to operate robots from multiple manufacturers and across different form factors. GRID onboarding includes building a digital representation of each robot, enabling simulation and training, adapting models, and providing the interfaces needed to control it. General Robotics works with OEMs as partners to make their robots compatible with the platform.

4. 23:57

   Three layers of robot training

   Kapoor divides training into general-purpose models, robotics-focused foundation models, and post-training for a particular robot and business task. He expects the first two layers to converge as broader models gain more physical-world understanding, while data will remain important for adapting robots to specific deployments. He argues that post-training is closest to customer value and may therefore capture more of the economics.