Google Cloud launches Gemini Robotics ER 2 to power physical AI startups

The rise of physical AI

The robotics industry is currently experiencing a significant surge in investment and technological development. The category saw a jump from $10 billion in 2024 to a projected $50 billion by mid-September 2026. This growth is driven by breakthroughs in multimodal reasoning, high-speed physics simulation, and more affordable sensors.

Venture capital is moving toward physical AI because these hardware-integrated businesses create stable, long-term portfolios. Some analysts predict that the physical AI market could reach $50 trillion over the next decade. These trends are making it more feasible to deploy robots across a wider variety of real-world tasks.

Google Cloud is positioning itself as a primary partner for this shift by offering a unified development stack. The company leverages its experience with Google DeepMind and Waymo to help startups navigate the unique challenges of physical deployment. This infrastructure helps companies manage the high costs of compute-intensive workloads like physics simulation and model training.

A new two-tier architecture

Modern robotics is shifting toward a two-tier architecture to balance intelligence with real-time performance. In this setup, heavy vision-language-action (VLA) models are trained in the cloud but separated during actual deployment. This allows for a clear distinction between high-level planning and low-level execution.

Frontier models like Gemini Robotics ER 2 run in the cloud to handle spatial reasoning and video progress tracking. These models act as the

Scaling robotics-as-a-service

Startups offering robotics-as-a-service (RaaS) face a difficult balance between recurring subscription revenue and the high cost of compute-intensive workloads. Because physical AI tasks like physics simulation and fleet inference require massive power, rigid infrastructure can become a financial burden. To solve this, founders are using dynamic capacity management on Google Cloud to maintain predictable costs.

These developers use tools like the Dynamic Workload Scheduler to organize planned training runs and automate multi-shape fallback lists. They also utilize granular accelerator slicing on Google Kubernetes Engine (GKE) to ensure they are not paying for underutilized hardware. These technical strategies allow robotics companies to scale their fleets while keeping their operational expenses manageable as they grow.

Innovations in training and sensing

Recent industry trends highlight three specific shifts in how robotics companies develop their systems. Founders are now creating rich synthetic training environments using photorealistic real-to-sim world models to prepare robots for reality. Additionally, they are finding ways to adapt general policies to different robot bodies in just a few hours, which speeds up the development of diverse hardware.

Another key shift involves pairing tactile sensing with video failure detection so machines can self-correct during a task. Google's models also play a specific role in data processing for these systems. For example, Gemini Robotics-ER can turn video clips into robot-ready instructions, while Gemini 3 labels objects frame-by-frame. Furthermore, Gemini Flash helps curate thousands of hours of factory footage to identify idle moments.

Diverse robot applications

The technology is being applied to a wide range of products, from educational tools to industrial machinery. For instance, Hands On Robotics created Pupper, an affordable robot dog that uses the Gemini Live API for real-time voice and multimodal dialogue. This robot can follow users, perform tricks, and hold natural conversations in dozens of languages while matching the user's chosen personality and tone.

In the industrial sector, Standard Bots is developing AI-native robotic arms designed to automate manufacturing. They handle the entire stack, including the hardware, proprietary software, and AI, to serve everyone from family-owned machine shops to Fortune 500 enterprises. These diverse applications show how the new cloud-to-edge infrastructure supports everything from student research to heavy-duty factory automation.