The GPT Moment for Robotics Is Here

In this talk

1. 1:01

   A general model for many robots

   Wang describes a goal of building a model that can control different robots across tasks. He traces progress from language models for planning, through PaLM-E and RT-2, to Open X-Embodiment, where a model trained across 10 robot platforms performed 50% better than specialists in the comparison he describes.

2. 9:56

   The data and deployment challenge

   Wang says robotics lacks the internet-scale training data available for language models, and that data needs to be captured in a form models can use. He argues that training across varied robots can help handle hardware changes and enable transfer, including tasks that previously required hundreds of hours of collection. Open X-Embodiment helped make data available, but did not solve reproducible evaluation.

3. 15:05

   Mixed autonomy in real operations

   Wang describes work with YC companies Weave and Ultra: folding unfamiliar laundry in a laundromat and packing customer orders in a warehouse. The Ultra operation ran for a full day with limited human intervention. He says the team reached a laundry system good enough for its task about two weeks after setting the goal, while emphasizing that robotics remains a systems problem.

4. 23:02

   Cloud inference and a founder playbook

   Wang says many Physical Intelligence robot evaluations use models hosted in the cloud. The system overlaps inference with robot movement and uses real-time chunking to keep successive action sequences smooth. For vertical robotics startups, he recommends understanding the workflow, choosing where a robot can make the biggest difference, keeping hardware and data collection scrappy, evaluating deployments, and reaching economic break-even before scaling.

5. 32:42

   Building an ecosystem around robotics

   Wang expects more robotics companies across different industries as hardware and startup costs fall, and says Physical Intelligence aims to enable them with cross-embodiment models and open-source weights. He also identifies support services such as teleoperation, data collection, and annotation as opportunities. For research, he describes an internal system that monitors training runs and can fix errors, improving compute utilization by about 50%.