Six robotics startups show how they build on Google's AI and cloud tools

A model launch and a customer parade

As Google Cloud reported in a post by Darren Mowry, who leads its startup and investor ecosystem work, the company launched Gemini Robotics Embodied Reasoning 2, or ER 2. It is available on the Gemini platform and in AI Studio. Google describes it as bringing DeepMind's multimodal reasoning and long-horizon task planning for robots to Google Cloud. In plain terms, it is a model that looks at a scene, understands an instruction and works out the steps of a long job.

The rest of the post is about the companies using these tools. The author says funding for physical AI and robotics jumped from $10 billion in 2024 to more than $50 billion by mid-September 2026. He also says some people predict the field could become a $50 trillion market over the next decade. That is a forecast, not a measurement, and it comes from a company that sells to this market.

Planning in the cloud, moving on the robot

The most useful idea in the piece is a two-tier design. Large reasoning models, such as ER 2, run in the cloud. They handle spatial reasoning, task planning and tracking progress in video. Smaller models, such as Gemma, run directly on the robot and handle real-time motor control. The cloud models direct the on-robot ones.

Startups also train heavier vision-language-action models, called VLAs, in the cloud. A VLA takes in images and instructions and outputs robot actions. The split makes sense because a robot cannot wait on a network connection when it is gripping an object, but it can wait a moment for a plan.

Paige Bailey of Google DeepMind put the moment this way: "intelligence is crossing the physical threshold." She also noted the practical hurdles: slippery surfaces, strict timing limits and hardware safety rules, all of which mean software mistakes have physical consequences.

Six companies, six different jobs

The post profiles startups doing quite different work.

• Apptronik builds humanoid robots, including one called Apollo. It has partnered with DeepMind to put Gemini Robotics models into Apollo, so the robot can follow spoken instructions and handle tasks like picking and packing.
• Deccan AI sells training data. It holds more than 45,000 hours of first-person, multi-camera and motion-capture recordings from homes and factories. It uses Gemini models to turn clips into instructions and to label hands and objects.
• Dyna Robotics trains its own robot foundation models on Google Cloud GPU clusters. It says it can ship updated models to production sites every day.
• Hands On Robotics makes Pupper, a build-it-yourself robot dog for teaching students. It uses the Gemini Live API for voice conversation.
• Simbe scans store shelves with robots and processes the results in the cloud.
• Standard Bots makes industrial robotic arms and runs its software on Google's AI Hypercomputer.

The post also says 15 European startups graduated from the first Google DeepMind Accelerator: Robotics cohort in London last month.

The money problem behind the machines

The piece raises a business point that is easy to miss. Enterprise buyers often want robots-as-a-service, a subscription model where the customer pays over time instead of buying outright. The author says this can strain a startup's cash flow. Meanwhile, simulation, training and running robot fleets all demand a lot of computing power, and idle capacity is costly.

Google's answer is flexible capacity: scheduling planned training runs and automatically falling back to other machine types when the first choice is unavailable. The post presents this as how founders keep costs predictable. It gives no figures on savings.

For the full pipeline, Google lists simulation tools, training systems and on-device software for factories and warehouses. It also names Intrinsic, a robotics software company in the Google family, as part of the stack.

How to read a vendor showcase

This is a customer showcase written by the company that sells the tools. It gives no independent measurements of how well any of these robots work. It also does not say how much each startup built itself versus relied on Google for, apart from a few cases such as Dyna's training setup.

Still, it is a useful map of where startups spend effort: data collection and labeling, simulation, cloud training and small on-robot models. Whether any of this holds up in a busy warehouse or store is a question the post does not answer.