Closing the Sim-to-Real Gap: Why AI Still Needs Real-World Data

Simulation promised to solve physical AI's data problem.

For some things, it has. For others, teams are finding out the hard way that it hasn't. We wanted to get people who are building physical AI in a room to chat about all about this: sim-to-real gap, robustness, and distribution shift.

Here's what we will cover:

• The sim-to-real gap in practice: where simulation held up, and where it fell apart on deployment
• Robustness under distribution shift: what actually breaks when a model meets a warehouse or road it wasn't trained for
• How much real data is actually enough: How teams decide when to stop trusting the simulator
• Where sim earns its keep: the cases where it's the right call and why that's not the whole story

About the panel:

• David Watkins: David Watkins is Research Lead at Tutor Intelligence, a robotics startup in Watertown, MA building intelligent robots for supply chain and logistics. He holds a PhD from Columbia University in learning mobile manipulation and has spent about a decade in robot manipulation research.
• Aravind Kandiah: Aravind is CTO and Co-founde Bifrost, which recreates the world in simulation so robotics teams can test and improve physical AI at the speed of compute, from Mars exploration with NASA to automating high-risk industrial work.

Register to attend live or receive the recording.