1 min read
Add as a preferred source on Google

Nvidia Robotics Chief Says OpenAI’s Astra Can Speed Robot Training Data 100-Fold

Deepu Talla said Astra can automate simulated scene creation; researchers may have needed months to build virtual objects and work settings in the past.

A robotic gripper lifts a steel hinge in a robotics lab / TokenPost.ai
A robotic gripper lifts a steel hinge in a robotics lab / TokenPost.ai

Nvidia robotics chief Deepu Talla said OpenAI’s Astra can generate some robot AI training data 100 times faster and automate simulated scene creation for researchers training models in virtual environments.

Talla said researchers may have needed months in the past to build virtual versions of objects and work settings a robot might encounter. Astra can accelerate that process, he said, and use engineering software to model mechanical parts. He also said it can control a robotic arm better than some specialized robotics AI models.

Robot AI models trained in simulation and released over the next six to 12 months may be more accurate than current leading models, Talla said. He predicted a robotics “ChatGPT moment” could happen within one to three years and said he believes general-purpose robots will need hands capable of dexterous manipulation.

OpenAI has also distributed robotics work across its research and Applied teams, a change covered in its robotics team restructuring.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…