Google launches Gemini Robotics 2 with whole‑body AI control

Google unveiled Gemini Robotics 2, its newest AI layer that lets robots plan and move from toe to fingertip, with early demos showing up to 92% success on specific dexterity tasks and a 14.7‑point gain in progress‑tracking accuracy over the prior model [1].

Whole‑body control and dexterity gains

Gemini Robotics 2 expands the vision‑language‑action (VLA) capability from tabletop dual‑arm work to full‑body motions such as walking, crouching, and reaching, demonstrated on the Apptronik Apollo 2 platform [1]. The system’s success rates vary by task: picking objects from a table (68.4%), from the floor (45.7%) and from a shelf (76.3%). Hand‑level actions with the 22‑DOF SharpaWave hand ranged from 32% (dustpan handling) to 92% (light‑bulb removal) [1]. These figures illustrate progress but also highlight that performance is still task‑specific and far from universal “human‑level dexterity.”

The reasoning model, ER 2, now tracks task progress in five video stages, achieving 57.4% accuracy—up from 42.7% in ER 1.6—and correctly identifies key video moments 91.3% of the time with sub‑second precision [1]. When directing a robot VLA, ER 2’s overall success rose to 60%, surpassing the 48.6% rate of its predecessor [1]. This improvement reflects the split‑labor architecture first used in generation 1.5, where high‑level planning (ER) and low‑level motion (VLA) operate separately, but now with broader whole‑body scope.

Safety, privacy and deployment limits

Safety testing shows ER 2 can detect a person within one metre with 93% accuracy and a mean error of 0.36 m, a marked improvement over ER 1.6’s 51% accuracy and 1.40 m error [1]. However, DeepMind’s safety report notes that when false‑stop rates stay below 5%, missed detections still exceed 40%, and reducing misses raises unnecessary stops to 15‑25% [1]. The model card explicitly warns against use in high‑risk domains such as medicine or transportation, and developers must handle privacy—obtaining consent and minimizing captured data—when robots operate near people [1].

Market context and competitive outlook

Google positions Gemini Robotics 2 as a step toward “intelligent whole‑body control” that can adapt to new robot bodies in a few hours, a claim made in its own blog [2]. Competitors like Boston Dynamics and Agility Robotics have focused on pre‑programmed locomotion and limited manipulation; Gemini’s combined VLA and ER architecture aims to reduce the engineering effort required to transfer skills across platforms. Yet, the public demos involve only two specific robot models, leaving broader fleet interoperability unproven [1].

What to watch

• Private preview rollout – Availability of the VLA model for early‑access partners and the timeline for on‑device version release.
• Safety certifications – Whether Google pursues functional‑safety standards (e.g., ISO 10218) to enable use in regulated environments.
• Competitor responses – Updates from firms such as Boston Dynamics or Agility Robotics that could narrow the whole‑body capability gap.

Google’s Gemini Robotics 2 demonstrates a shift from isolated arm dexterity toward integrated whole‑body reasoning, but real‑world impact will hinge on consistent task success across diverse robots, safety validation, and broader ecosystem adoption.