Whole-body motor control arrives in physical AI via Gemini Robotics 2

The Takeaway

• A Vision-Language-Action (VLA) model converts visual input into direct motor actions across entire humanoid bodies.
• Gemini Robotics ER 2 coordinates multi-robot teams while evaluating task progress across five distinct tiers.
• Local adaptation for new bi-arm setups requires fewer than 200 demonstration examples over a few hours.
• A safety benchmark called ASIMOV-Agentic measures when systems should refuse unsafe tool commands.

Integrating physical coordination with AI reasoning requires moving beyond basic tabletop actions. Google DeepMind introduced Gemini Robotics 2 to serve as an intelligent motor and reasoning layer for physical hardware. The August 2026 update establishes three distinct tools: the primary Vision-Language-Action (VLA) model for physical movements, the Gemini Robotics ER 2 Vision-Language Model (VLM) for agentic reasoning, and Gemini Robotics On-Device 2 for local execution.

Whole-body motor control allows humanoid machines to crouch, stretch, balance, and walk inside tight physical spaces. Operating the Apptronik Apollo 2 humanoid, the primary model translates natural language requests into complex multi-step actions, such as fetching a watering can from a table and placing it on a lower shelf. Finesse extends down to the hands. The system operates 5-fingered, 22 degree-of-freedom SharpaWave hands for delicate tasks like tying knots or sealing ziplock bags, alongside standard 2-fingered parallel grippers on Franka Duo platforms. When operating locally, the on-device variant adapts to new physical hardware like Dexmate, SO101, and Trossen platforms in a matter of hours.

Converting visual inputs and spoken instructions directly into full-body motor control across diverse physical bodies represents the primary functional shift. A single model checkpoint controls entirely different hardware setups without needing full retrain cycles for each body shape.

Gemini Robotics 2 platform specifications

Developers across India can access the reasoning model through Google AI Studio or generate keys via the standard API portal. Enterprise teams requiring private infrastructure can test preview access on Vertex AI. You might already know how to Put your family in AI art with Gemini India using consumer software. This expansion brings that same core artificial intelligence stack to physical enterprise hardware. Full end-to-end physical execution in local facilities remains tied to when hardware partners ship commercial units into the Indian market. Lower-level motor controls remain limited to early-access partners for now.

The Unboxed Truth

This software update brings physical AI much closer to managing actual working environments. At Unbox Daily HQ, we see this leap from tabletop upper-body movement to full-body coordination as a crucial milestone for real automation. At $2.00 (approx ₹168) per million input tokens, developer API testing costs less than a cup of specialty coffee, keeping software testing cheap enough for local teams to build complex workflows immediately. You cannot deploy these full-body controls on a local assembly floor tomorrow because commercial humanoid bodies are not yet sold locally. You can, however, start building and simulating multi-step agentic pipelines today.

Best for: Engineering teams testing autonomous robotics software and multi-machine coordination.

Who Is This For: Software engineers and robotics researchers aged 25 to 45.

Courtesy: Google DeepMind