Google DeepMind has officially launched its first robotics accelerator, and the inaugural cohort is already in London. The program is a three-month initiative for early-stage robotics startups across Europe, with the selected founders gathering this week to kick off the program and meet the Google DeepMind and Google teams. The 15 companies span nine countries and cover everything from brain-navigating microrobots to electronic skin for robot arms , and they're all getting direct access to Google's most advanced AI infrastructure.

What's actually on the table

The program is equity-free. That detail matters: startups receive mentorship, infrastructure access, and product guidance without surrendering any ownership stake. That's a meaningful distinction from traditional corporate accelerators that often take 5–10% equity in exchange for support.

The concrete benefits include:

• Access to Google's full AI stack, technical expertise, and Gemini robotics models.
• Tailored training for individual technical and business challenges, plus exclusive invitations to tech bootcamps hosted by Google, with dedicated mentoring from Google teams and AI experts.
• Direct engagement with Google AI and Gemini Robotics models to help move startups from prompt to production.
• Eligibility for up to $350,000 in credits from the Google for Startups Cloud program.
• Ongoing support through the Google Accelerator Alumni Network after graduation.

The program begins with a five-day in-person event in London, followed by online training and mentoring sessions through September, and a three-day graduation event later that month. It will culminate in a Demo Day in London.

The Gemini Robotics stack, explained

The AI backbone these startups are getting access to is not generic. Google DeepMind introduced a new family of AI models purposefully designed for robotics and built upon the foundation of Gemini 2.0, with Gemini Robotics being an advanced Vision-Language-Action (VLA) generalist model capable of directly controlling robots. A VLA model is one that takes in visual input and natural language instructions and outputs physical motor commands , bridging the gap between language models and hardware.