On July 30, 2026 — two days ago — Google DeepMind published something that the robotics world had been waiting for without quite knowing it.
Not a new robot. Not a new chip. Not a new hardware product you can buy.
Something more fundamental than any of those things. A brain.
Gemini Robotics ER 2 is Google DeepMind's answer to the problem that has quietly limited every humanoid robot ever built: they can move, but they cannot think across time. They execute the task in front of them. They do not understand the broader mission. They cannot recover when something goes wrong mid-sequence. They cannot coordinate with another robot to share the work.
ER 2 changes all of that. And it did it in one announcement that most people outside the robotics world barely noticed.
Before Gemini Robotics ER 2 — The Problem It Solved
To understand why this matters, you need to understand what robots could and could not do before July 30, 2026.
Before — What Robots Were Good At
Robots before ER 2 were extraordinarily good at specific, defined, repeatable tasks. Pick this object from this position and place it there. Weld this seam. Sort these packages by size. Spray paint this surface.
The key word is specific. Pre-programme the task precisely. Build the environment around the robot. Control every variable. When all of those conditions are met, industrial robots are faster, more precise, and more consistent than any human.
Before — What Robots Were Terrible At
Everything that requires thinking across multiple steps, adapting to variation, or recovering from failure.
Earlier Gemini robotics models focused mainly on upper-body and tabletop tasks. The robot could control its hands. It could not control its whole body. It could perform a task on a table in front of it. It could not walk to where the task was, perform it, then move to the next task.
More critically — if something went wrong in the middle of a multi-step task, earlier systems had no reliable way to recover. If step 3 of a 10-step sequence failed, the robot typically had to restart the entire sequence from the beginning. Every failed attempt wasted time. Every restart required human intervention to reset the environment.
And robots could not work together meaningfully. Two robots in the same space were essentially operating independently, with no shared understanding of the overall task or each other's progress.
This was the state of physical AI on July 29, 2026.
What is Gemini Robotics ER 2 — The Simple Explanation
Gemini Robotics ER 2 is a high-level brain for robots. It allows robots to chat with humans, understand the physical world, and plan multi-step tasks. It then hands off motor execution to any given lower level vision-language-action (VLA) model.
The word "brain" here is not metaphor. It is a precise description of the architecture.
Think about how humans handle a complex task — say, tidying a room. You do not just execute a single programmed movement. You assess the room. You make a plan. You start executing. You notice something unexpected — the bin is full. You adapt — empty the bin first. You continue. You check your progress. You know when you are done.
ER 2 gives robots that entire capability — assess, plan, execute, monitor, adapt, know when done — as a high-level reasoning layer that sits above the physical movement systems.
Gemini Robotics ER 2 is based on Gemini 3.5 Flash. It supports text, image, video, and audio inputs with a context window of up to 128,000 tokens and text output up to 64,000 tokens.
This is built on the same underlying AI model that powers Gemini 3.5 Flash — one of Google's most capable and efficient language models — extended with what Google calls embodied reasoning. The ability to understand physical spaces, objects in those spaces, and how to interact with them.
The Three Models in Gemini Robotics 2 — Explained Simply
Google DeepMind launched three models simultaneously on July 30. Understanding how they relate to each other is important.
Model 1 — Gemini Robotics ER 2 (The Brain)
Gemini Robotics ER 2: Our most capable embodied reasoning model. It is a vision language model that acts as our agent, enabling robots to communicate with humans, understand the physical world and plan multi-step tasks lasting several minutes.
ER 2 is the planning and reasoning layer. It receives instructions from humans, understands the physical environment through cameras and sensors, breaks the task into steps, monitors execution, and recovers when things go wrong. It does not directly control the robot's motors. It tells the lower-level models what needs to happen.
Model 2 — Gemini Robotics 2 (The Body Controller)
Gemini Robotics 2, a vision-language-action (VLA) model that converts visual and language inputs into robot movements and can control full humanoid robots as well as dual-arm systems. Unlike certain earlier models, it can control all of a humanoid robot's components and not just its hands. As a result, the algorithm can optimize the host machine's centre of gravity in a way that minimises the risk of falls.
While earlier Gemini robotics models controlled the humanoid's upper-body to achieve table-top tasks, Gemini Robotics 2 expands physical AI into whole-body motions. For the first time, our model can now control entire humanoid robots, translating intent into intelligent whole-body control.
In a demonstration, an Apptronik Apollo 2 humanoid was asked to place a watering can in a green bin on a lower shelf. The robot walked to the object, picked it up, moved to the shelving area, and placed it in the specified location. That required whole-body coordination — walking, bending, reaching, grasping, moving, placing — all controlled by a single AI model.
Model 3 — Gemini Robotics On-Device 2 (The Edge Model)
Gemini Robotics On-Device 2, an optimised VLA model that runs locally on robotic hardware. To support industrial deployments where cloud connectivity is limited, this model runs directly on robot hardware.
The model can adapt to new two-arm robot designs with fewer than 200 examples and several hours of training, including platforms with different shapes, sensors and ranges of motion.
Fewer than 200 training examples to adapt to a completely new robot. A few hours of training. This is remarkable. Previously adapting an AI model to new robot hardware required thousands of examples and weeks of training. On-Device 2 does it in hours with minimal data.
The Five Breakthrough Capabilities
Breakthrough 1 — Whole Body Control
Before ER 2, Gemini robotics controlled robots from the waist up. After ER 2, the model controls the complete robot — feet, legs, torso, arms, hands, and head simultaneously. Walking, bending, balancing, reaching — all coordinated as a single unified system rather than separate subsystems.
Breakthrough 2 — Task Recovery Without Restarting
If the robot makes a mistake, ER 2 can identify the last step that the machine completed correctly and pick up where it left off. That removes the need to redo chores from scratch, which saves time.
This sounds simple. It is genuinely revolutionary for practical deployment. A robot that can restart from failure without human intervention is a robot that can work unsupervised. A robot that must restart from the beginning every time something goes wrong requires constant human oversight to reset the environment.
Task recovery without full restart is what separates a useful robot from an expensive demo.
Breakthrough 3 — Knowing When a Task is Complete
Gemini Robotics ER 2 now understands when tasks begin and end, and can pinpoint the moment key events occur, marking a step change in progress understanding.
This seems obvious but was genuinely difficult. Knowing definitively that a task is complete — that the cup is properly placed, the package is correctly sealed, the shelf is fully stocked — requires understanding of what "done" means contextually, not just checking a position coordinate.
Breakthrough 4 — Multi-Robot Collaboration
Multiple robots can communicate, recognise each other's unique physical strengths, and autonomously delegate tasks to complete a shared mission.
Gemini Robotics ER 2 allows different robot types to communicate and coordinate on complex workflows that would otherwise require a single machine to perform every task.
A large robot and a small robot working in the same warehouse. ER 2 understands that the large robot can reach high shelves and carry heavy loads. The small robot can navigate narrow aisles. ER 2 coordinates them — automatically delegating tasks based on each robot's physical capabilities — without a human planning the division of labour.
This is genuinely new. Before ER 2, multi-robot coordination required explicit programming of every handoff. Now the AI handles coordination autonomously.
Breakthrough 5 — Think and Act Simultaneously
The design of Gemini Robotics ER 2 allows the robot to "think" about what comes next while simultaneously performing its actions.
Before this, robots executed a step, then paused to plan the next step, then executed, then planned. Every pause added latency and made the robot appear hesitant and slow. ER 2 plans the next step while executing the current one — creating fluid, continuous motion that looks and feels far more natural.
The Honest Performance Numbers
Google published actual performance data rather than just demo videos. Here is what the technology can and cannot do right now:
Tests with Apollo 2 showed a 68.4% success rate when picking up objects from a table, 45.7% from the floor and 76.3% from a shelf. Google DeepMind also said the robot movement still needs to become faster and more reliable.
Tests showed stronger performance with two-finger grippers than multi-finger hands.
This is honest engineering. 68.4% success on table pickup. 45.7% from the floor. Not 100%. Not even close to 100%. Google published these numbers rather than hiding them — which is the mark of genuine scientific rigour rather than pure marketing.
The direction matters more than the current numbers. These are first-generation whole-body numbers. The improvement curve for AI systems has been consistent — significant gains every six to twelve months. A 45.7% floor pickup rate in July 2026 will be 65% by early 2027 and 80% by late 2027 if the pattern holds.
Where Can You Access Gemini Robotics ER 2 Right Now
Gemini Robotics ER 2, our reasoning model, is now available on Google AI Studio and in private preview on Gemini Enterprise Agent Platform. Our VLA and On-Device models are available to early-access partners.
Available now:
✅ Gemini Robotics ER 2 → Google AI Studio
✅ ER 2 → Gemini Enterprise Agent Platform
(private preview)
Early access partners only:
⚙️ Gemini Robotics 2 (VLA model)
⚙️ Gemini Robotics On-Device 2
Hardware demonstrated on:
→ Apptronik Apollo 2 (with SharpaWave hands)
→ Apptronik Apollo 2 (with Inspire hands)
→ Franka Duo with Robotiq gripper
Before and After — The Complete Comparison
| Capability | Before ER 2 | After ER 2 |
|---|---|---|
| Body control | Upper body only | Full whole-body ✅ |
| Multi-step tasks | Limited | Minutes-long sequences ✅ |
| Task recovery | Restart from beginning | Resume from last correct step ✅ |
| Task completion | Difficult to detect | Knows when done ✅ |
| Multi-robot | No coordination | Autonomous delegation ✅ |
| Think while acting | Sequential only | Simultaneous ✅ |
| New robot adaptation | Weeks + thousands of examples | Hours + <200 examples ✅ |
| Offline operation | Cloud required | On-device available ✅ |
| Human communication | Limited | Natural language chat ✅ |
What This Means for the Future of Robotics
Gemini Robotics ER 2 is not a product. It is infrastructure. It is the intelligence layer that any robot manufacturer can build on top of.
Apptronik already demonstrated it on the Apollo 2. The Franka Duo demonstrated it with a robotic gripper. The model works across different hardware because it is designed to work across different hardware — adapting to new robot embodiments in hours rather than months.
Google DeepMind describes Gemini Robotics 2 as the "intelligence layer" for robots, enabling machines to move beyond pre-programmed tasks and adapt to changing environments. The company says the platform is designed to work across different robot types and can be adapted to new hardware in a matter of hours.
The implication is significant. Every robot manufacturer who partners with Google gets access to the most capable AI brain available for physical robots. The hardware competition between robot manufacturers continues. But the intelligence layer — the reasoning, planning, recovery, and coordination — could increasingly come from Google.
This is the same pattern that played out with smartphone operating systems. Many manufacturers made the hardware. Google's Android provided the intelligence layer. The result was that Google became central to the smartphone ecosystem without manufacturing a single phone.
Gemini Robotics ER 2 is the beginning of the same strategy applied to the physical world.
TheTechGenAI Studios — We Cover the AI That Changes Everything
At TheTechGenAI Studios, we track every major AI announcement that reshapes what is possible — from Gemini Robotics ER 2 changing what robots can think, to AI-powered websites that change how your business finds customers.
Talk to us about your project. Free consultation, no commitment.
📧 [email protected] 📱 +91 98704 05692 🌐 thetechgenai.com
FAQs
What is Gemini Robotics ER 2? Gemini Robotics ER 2 is an embodied reasoning AI model launched by Google DeepMind on July 30, 2026. It acts as the high-level brain for robots — planning multi-step tasks, communicating with humans, recovering from failures, and coordinating multiple robots. It is based on Gemini 3.5 Flash with a 128,000 token context window.
What is the difference between Gemini Robotics ER 2 and Gemini Robotics 2? ER 2 is the reasoning and planning layer — the brain. It understands instructions, plans tasks, monitors progress, and coordinates robots. Gemini Robotics 2 is the body control layer — it translates ER 2's plans into actual robot movements including whole-body control of humanoid robots.
What changed before and after Gemini Robotics ER 2? Before: robots controlled upper body only, restarted entirely on failure, could not coordinate with other robots, planned sequentially. After: whole-body control, resumes from last correct step on failure, autonomous multi-robot coordination, plans next step while executing current one.
What are the current success rates for Gemini Robotics 2? Google published honest numbers: 68.4% success picking from a table, 45.7% from the floor, 76.3% from a shelf. Google acknowledged the robot still needs to become faster and more reliable.
Where can I access Gemini Robotics ER 2? Gemini Robotics ER 2 is available on Google AI Studio now, and in private preview on Gemini Enterprise Agent Platform. The VLA and On-Device models are available to early-access partners only.
What robots does Gemini Robotics 2 work with? It has been demonstrated on the Apptronik Apollo 2 (with multiple hand configurations) and Franka Duo with Robotiq gripper. Gemini Robotics On-Device 2 can adapt to new robot hardware with fewer than 200 training examples and a few hours of training.
