One Brain, Any Body: Google DeepMind's Keerthana on Gemini Robotics 2, Cross-Embodiment & Humanoids
Keerthana Gopalakrishnan, staff research scientist and Gemini Robotics research lead at Google DeepMind, discusses robotics progress, general-purpose robot models, and humanoid deployment with the host of The Cognitive Revolution.
In this talk
- 4:33
What running robots show
Keerthana says humanoid running is an impressive public benchmark, but speed is not usually what makes robots useful. Locomotion benefits from simulation because flat surfaces and predictable contact are easier to model. Manipulation is harder when objects deform or have complex friction, as with cloth.
- 13:42
Why robotics is still early
Keerthana sees generality and task mastery as distinct goals: a broad foundation can make it cheaper to adapt a robot to many tasks. She places robotics around the GPT-2 stage because few-shot learning and performance across different robot bodies remain limited. Demonstrations can help, but results depend on the task and whether the model can adapt when the scene changes.
- 24:20
How the Gemini Robotics models fit together
Gemini Robotics-ER handles higher-level reasoning and tool use, while Gemini Robotics translates intentions into robot actions; the smaller On-Device model runs on the robot’s computer. ER is available more widely through an API, while the action models are being developed with trusted testers and partners. Keerthana says the VLA can control a humanoid from fingertips to feet, with other decisions handling stabilization.
- 51:02
Deployment depends on reliability and safety
Keerthana says cross-embodiment progress is real, but reliably transferring a model to a new robot remains difficult. Errors can compound across multi-step tasks, and retries make tasks such as Lego assembly easier than frying an egg, where a mistake creates a mess. Robot safety includes physical stability, responding to sensor failures, and following human instructions.
- 1:13:35
Data and robot interaction will vary
Keerthana says there is no settled answer on whether robotics needs predictive world models; the field should test different approaches. She expects a mix of teleoperation data, wearable-device data, and human video because each offers different tradeoffs in scale and precision. People may need both simple language instructions and more detailed control, especially for personal preferences and safety.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.