Google Gemini Robotics 2 adds whole‑body control and multi‑robot

Google unveiled Gemini Robotics 2, a vision‑language‑action model that lets humanoid robots perform whole‑body movements, fine‑grip manipulation and coordinate with other robots, marking a step toward adaptable, general‑purpose physical AI【1】. The advance matters because it moves robots beyond pre‑programmed, single‑task routines toward autonomous operation in dynamic, real‑world settings.

Whole‑body intelligence and dexterity

Gemini Robotics 2 expands on earlier models that handled only upper‑body tasks by controlling a robot from feet to fingertips. In a live demo, the model guided Apptronik’s Apollo 2 humanoid to walk, crouch, stretch and place a watering can on a lower shelf, showing the ability to translate natural‑language commands into coordinated full‑body motion【1】【2】. The same checkpoint also operated on the Franka Duo platform, proving cross‑embodiment flexibility.

Dexterity improvements include controlling Apollo 2’s 22‑degree‑of‑freedom five‑fingered hand to tie knots, seal zip‑lock bags and unscrew light bulbs, while two‑finger grippers on Franka Duo performed precise insertion and packing tasks【1】【2】. Google notes ongoing work to reach human‑level speed and precision, but the current level already surpasses prior tabletop‑only manipulation.

Reasoning, teamwork and on‑device adaptation

The companion model Gemini Robotics ER 2 serves as a high‑level reasoning engine, breaking spoken instructions into multi‑step plans, monitoring progress and handling failures, enabling tasks that last several minutes and involve hundreds of decisions【1】【2】. A new multi‑robot collaboration feature lets different robots communicate and share workloads, a capability absent from most existing systems.

Gemini Robotics On‑Device 2 is optimized for local execution, allowing robots to adapt to new embodiments with fewer than 200 examples over a few hours, eliminating reliance on cloud connectivity and reducing latency【1】. This fast adaptation was demonstrated on diverse platforms such as Dexmate, SO101 and Trossen.

Competitive context

Compared with conventional industrial robots that require extensive reprogramming for each new task, Gemini’s models aim for rapid skill transfer across bodies, a claim Google attributes to its “motion transfer” techniques inherited from Gemini 1.5【1】. While other firms like Boston Dynamics and NVIDIA have showcased locomotion and manipulation, Google’s integration of vision, language and action in a single model, plus on‑device operation, differentiates its approach. No independent benchmark data are provided, so performance relative to rivals remains unverified.

What to watch

• Rollout timeline – Google has released the models to early‑access partners; monitoring broader commercial availability will indicate market impact.
• Safety benchmarks – Google introduced the ASIMOV‑Agentic safety benchmark to evaluate unsafe tool calls and human proximity handling; future test results will gauge real‑world safety compliance.
• Competitor responses – Expect announcements from other AI‑robotics players (e.g., NVIDIA, Boston Dynamics) addressing whole‑body control or on‑device inference as the field evolves.

Google’s Gemini Robotics 2 demonstrates that large‑scale multimodal AI can now drive full‑body robot behavior, dexterous manipulation and collaborative workflows, narrowing the gap between fixed‑function automation and adaptable physical agents. The open question is whether the rapid adaptation and safety mechanisms will translate into reliable, large‑scale deployments across homes and factories.