Maintain real-time inference speed for high-capacity generalist robotic manipulation policies

Develop algorithms and system designs that preserve real-time inference speed while scaling the capacity of generalist robotic manipulation policies, ensuring responsiveness during deployment.

Background

Generalist policies often use large-capacity architectures that can impair real-time control. The authors explicitly cite maintaining real-time inference speed as an open research problem even as larger models demonstrate improved capabilities.

Solving this will enable practical deployment where responsiveness and safety are critical.