From a moving image to a useful action.

2 stories · 2 minute read

01 / CONVERSATIONAL AI / COMPANY ANNOUNCEMENT

Gemini adds a face while tools work in the background

Google’s September 24 Live Avatar release pairs generated video with Gemini 3.8 Live speech in Gemini Enterprise. Agents can call tools while continuing the conversation. Preset avatars are available; creating a custom avatar still requires enterprise allowlisting.

Read the source ↗ blog.google

02 / ROBOTICS / RESEARCH INTERVIEW

A robot trajectory is only the start of a policy

Conceptual pipeline from human video to robot trajectory, followed by a separate learning step for a reactive policy. — Conceptual illustration based on the researchers’ interview and project description. Reconstructing and replaying a motion does not by itself teach recovery when the scene changes.

In a new RoboPapers interview, the Do As I Do researchers explain how ordinary videos become executable robot-hand trajectories. Those open-loop replays supply training data; handling changed objects or positions still requires learning a robust, observation-driven policy.

Read the source ↗ robopapers.substack.com