In 30 seconds
Google DeepMind presents a brief demonstration involving a spoken request to erase a whiteboard. The clip is a single example, not a comprehensive evaluation.
- 1A spoken instructionA person asks the robot to erase a whiteboard.
- 2A physical taskThe robot acknowledges the request in the demo.
- 3Limits of a demoOne clip does not measure reliability across tasks.
Takeaways
- A person gives a robot a spoken task.
- The robot acknowledges the request.
- A short demonstration does not establish general reliability.
Read the edited text by chapter
The request
A person asks a robot to erase a whiteboard. Google DeepMind presents the short clip as a demonstration of Gemini Robotics on the ALOHA 2 robot.
A person asks a robot to erase a whiteboard. Google DeepMind presents the short clip as a demonstration of Gemini Robotics on the ALOHA 2 robot.
The demonstration
The robot responds that it will erase the board. The demonstration illustrates an instruction and a physical task, but the brief clip does not establish how reliably the system works across other settings.
The robot responds that it will erase the board. The demonstration illustrates an instruction and a physical task, but the brief clip does not establish how reliably the system works across other settings.
A single example
At the end, the person praises the robot. The video provides a single example rather than an evaluation of performance, safety or general-purpose capability.
At the end, the person praises the robot. The video provides a single example rather than an evaluation of performance, safety or general-purpose capability.
More from this channel
38:10EN
Google DeepMind· Oct 1, 2026From deepfakes to DNA: how SynthID watermarks AIHannah Fry, Pushmeet Kohli, Jeremy Ratcliffe · Google DeepMindHannah Fry sits down with Pushmeet Kohli and Jeremy Ratcliffe to explain SynthID: a mathematical watermark now used on text, images and video, and being extended to protein sequences so DNA synthesis companies can tell AI-designed molecules from nature.
44:25ENGoogle DeepMind· Sep 9, 2026How AI Is Changing Weather ForecastingHannah Fry, Peter Battaglia · Google DeepMindHannah Fry and Peter Battaglia discuss hurricane forecasts, uncertainty and WeatherNext 3.
44:41ENGoogle DeepMind· Aug 26, 2026The Mathematics of AI UncertaintyZoubin GhahramaniZoubin Ghahramani explains Bayesian updating, calibration and their computational challenges.
10:52PT → EnglishGoogle DeepMind· Jul 14, 2026Reconstructing Pelé’s Lost GoalHistorical research, actors and AI recreate a 1959 play without original footage.
53:05ENGoogle DeepMind· Jul 10, 2026Inside AI Models: Neel Nanda on Interpretability and SafetyHannah Fry, Neel Nanda · Google DeepMindHannah Fry explores reasoning traces, probes and sparse autoencoders.
42:38ENGoogle DeepMind· Jun 23, 2026When AI Agents Delegate: Security and a Possible Agent EconomyHannah Fry, Nenad Tomašev · Google DeepMindNenad Tomašev discusses specialists, oversight and correlated failures with Hannah Fry.