Gemini Robotics ER 2 (The Gemini Embodied Reasoning (ER) model)— It’s an intelligence layer- not a robot. It’s demonstrated its performance on robots mentioned below. It makes robots think, act, and accomplish. Three major achievements are:
- 1- Intelligent whole-body control — walk in a house, building, or on roads to accomplish assigned tasks. Gemini Robotics 2 can manage whole-body, hand, and other human-like movements as required. It does this by managing the center of gravity.
- 2- Advanced dexterity — clean a room, assemble parts or gadgets in a complex box; this is a big leap, as it would be highly useful in real-life scenarios.
- 3- Multi-robot collaboration — talk with other robots to get things done.
Google DeepMind provides the core intelligence, while Gemini Robotics implements the robotics features.
Gemini Robotics 2 can control many kinds of robots to accomplish a physical-world task, as demonstrated here with:
- 1. Apptronik Apollo Robotics and Duo robots,
- 2. Apptronik’s Apollo with Inpire hands
- 3. Apptronik’s Apollo with Sharpa hands.
Apptronik, a Robotics company, partnered with Google DeepMind and Gemini Robotics to give the physical robot the essence of human-like intelligence. Apptronik’s Apollo 2 robot has years of experience and is one of the best in Apptronik’s robot crew. It is designed to work with humans and can improve with experience. It comes as a bipedal version for human-like places or as a wheeled frame for other generic applications.
The Franka Robotics F3 Duo has two highly dexterous mobile manipulation arms, vision cameras, and is not a humanoid robot. It can generate and collect rich data from its environment while executing tasks. Inpire Hands by Inpire Robotics, a Chinese company, are five-fingered dextrous hands. The Sharpa Hands are human-like robotic hands by a Singapore-based company, Sharpa.
Gemini Robotics On-Device 2 can work without an internet connection and is a real hit for real-time applications such as street surveillance at night, as these robots can have night-vision cameras installed. I suggest Gemini Robotics 2’s next release should also work with night-vision cameras. This can be helpful for gun violence crimes where robots can detect and stop such dangerous actions by just being installed on a pillar post.
Multi-robot collaboration — talk with other robots to get things done. Apptronik Apollo takes human instructions and converts them into mini instructions for Duo to perform with high accuracy. This has been demonstrated in their release.
Vision Language Action (VLA) — Google demonstrated VLA converted to motor control with Apptronik’s Apollo’s robots, showing it pick up a watering can, tie a bag, take a bulb out from a lamp, pack grapes for refrigeration, pick up a shopping bag, take bread out of a toaster, and with smaller robots’ dexterity being elaborately explained. New robots can be trained within hours now with this model.
Vision Language Model (VLM) — To talk to humans and other robots and understand the task to be accomplished and perform it. Example: making Duo do things as per what Apollo, backed by Gemini ER 3, wants it to.
Its performance on physical tasks is better than Gemini ER 1.6; its performance on progress classification tasks (It tracks robot performance toward task completion) is better than GPT 5.6 Sol, and Opus 5. For precision moment finding, which determines when to stop and when to start (for example, when to stop pouring tea into a mug), ER 2 is just behind Opus 5 and above GPT 5.6 Sol in this task. In success detection (both video- and image-based), it is above GPT 5.6 Sol and Opus 5. In question answering and generalized instrument-measuring data, it outperforms GPT 5.6 Sol and Opus 5, not to mention it’s better than Gemini ER 1.6.
Google DeepMind has added safety protocols to the model.
In summary, this is a big leap for Google DeepMind; the Gemini Embodied Reasoning model, (Gemini Robotics ER 2) is really interesting, while it leaves Gemini Robotics 1.5 behind. Still, there is a long way to go, for example, making Gemini ER Robotics models do night-vision work that humans can’t do. This could help law enforcement and space exploration when sunlight isn’t available. Maybe it is time for Google to sign up with space exploration robotics companies as well.
Well done, Google!
Thank you for reading.
Please subscribe for updates.
Artificial Intelligence and Future 