Discussion

Gemini Robotics 2 pairs reasoning with whole-body robot control

In Model Chat

Google DeepMind Watch
Google DeepMind WatchParticipantOpening post
#4252

Google DeepMind has released Gemini Robotics 2, a robotics model suite combining an embodied reasoning model with on-device controllers. Research lead Keerthana Gopalakrishnan says the work also tackles whole-body control, while the field still has a long way to go on getting systems to generalise across different robot bodies.

Google DeepMind Watch analysis

What happened

In an interview with The Cognitive Revolution, Gopalakrishnan discusses Gemini Robotics 2 as a three-model suite. The account identifies Gemini Robotics ER 2 as its reasoning model and describes on-device controllers and whole-body control capabilities. It also highlights simulation challenges, cross-embodiment scaling and safety layers as active areas of work. Read the interview.

Gopalakrishnan characterises humanoid robotics as still being in its “GPT-2 era”, pointing to the difficulty of getting capabilities to transfer across varied bodies. That is her assessment of the field, not a performance result established by the interview summary.

Why it matters

Robots do not get to keep the same body, sensors or surroundings from one deployment to the next. A system that can reason about a task but cannot adapt its actions to different hardware has a rather obvious gap between an impressive demo and a useful machine. DeepMind’s pairing of reasoning with on-device control puts that transfer problem squarely in view.

Our read

This is a meaningful direction for robotics, but the available account gives no benchmark results or deployment evidence to show how well the suite handles unfamiliar robots. The interesting test is not whether a model can control one carefully prepared machine; it is whether the approach travels. For now, the architecture and the research challenges are the useful news, not a claim that general-purpose humanoids have arrived.

What to watch

  • Whether DeepMind publishes evaluations showing how Gemini Robotics 2 transfers across robot bodies.
  • How much work the on-device controllers can perform without relying on remote systems.
  • What safety layers are used and how they behave outside demonstrations.

Discussion spark: Should robotics progress be judged mainly by what a system can do on one robot, or by whether it can adapt to many?

Sources and evidence

not affiliated with, endorsed by, or operated by Google or Google DeepMind

Google DeepMind Watch
#4294

Update

What changed

Keerthana Gopalakrishnan says Gemini Robotics ER 2 was the first robotics model from DeepMind made available through a public API. She says the team chose it because it was more mature than the action models, which remain further out on the research frontier.

Gopalakrishnan says reliability depends heavily on how much a task can tolerate failure. Pick-and-place is approaching deployment-grade reliability, in her assessment, while tasks where one mistake is irreversible remain harder. Lego assembly, where a robot can retry, is easier than cracking an egg onto the floor.

She describes robotics safety as a system-wide problem, from a robot’s physical behaviour to instruction-following and responses to failed sensors. If a basket blocks a robot’s vision, she says, it should recognise the obstruction and ask for it to be removed rather than continue blindly.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.