Google’s Gemini artificial intelligence is moving beyond screens and taking control of humanoid robots.

Google DeepMind has introduced Gemini Robotics 2, a new generation of AI models that allows robots to walk, crouch, handle objects, follow multi-step instructions, and work with other machines.

The development expands Gemini’s role in robotics from controlling upper-body movements to coordinating an entire humanoid, bringing Google closer to its goal of building general-purpose AI for the physical world.

Key facts at a glance

• Google DeepMind announced Gemini Robotics 2, a family of AI models for full-body humanoid robot control.
• The system extends control “from feet to fingertips,” moving beyond earlier upper-body-only robotics models.
• Gemini Robotics 2 comprises three models: Gemini Robotics 2, Gemini Robotics ER 2, and Gemini Robotics On-Device 2.
• In demonstrations, the system controlled Apptronik’s Apollo 2 humanoid for tasks such as placing objects, crouching, tying bags, and changing lightbulbs.
• Reliability remains mixed, with success rates ranging from 32% for using a dustpan to 92% for unscrewing a lightbulb.

Gemini gains full-body control

In demonstrations, Google showed its AI controlling several robots, including Apptronik’s Apollo 2 humanoid. One demonstration showed Apollo 2 responding to a request to place a watering can inside a green bin on a lower shelf. The robot walked toward the object, picked it up, moved to the shelf, crouched, and placed the watering can in the correct location. Previous versions of Gemini Robotics primarily controlled a humanoid’s upper body and focused on tabletop tasks. Google DeepMind said the new system extends that control “from feet to fingertips.”

Gemini Robotics 2 also improves the robot’s ability to use its hands. In other demonstrations, Apollo 2 tied a trash bag, sealed a plastic storage bag, operated a dustpan, and installed or removed a lightbulb. These tasks require precise force control, spatial awareness, and the ability to adapt to deformable objects. A trash bag, for example, changes shape as it is gathered, so the robot must adjust its grip in real time. Similarly, inserting a lightbulb requires careful alignment and just the right amount of torque. The fact that the system can attempt such jobs marks a clear step forward from earlier tabletop manipulation benchmarks.

Three AI models power the robots

Google divided the system into three models. Gemini Robotics 2 translates visual information and human instructions into physical movements. Gemini Robotics ER 2 acts as the planning system, allowing robots to understand their surroundings, divide assignments into smaller steps, and track their progress. Gemini Robotics On-Device 2 runs locally on robotic hardware, allowing machines to operate without a constant internet connection. Google said the model can adapt to a new two-armed robot using fewer than 200 training examples collected over several hours. This on-device capability is important for real-world deployments in warehouses and factories, where network connectivity can be unreliable and latency must be kept to a minimum.

Gemini Robotics ER 2 also introduces multi-robot collaboration. This allows different machines to communicate and divide a larger assignment between them, potentially supporting more complicated workflows in factories and warehouses. In such a setup, one robot might fetch components while another assembles them, and a third might package the finished product. The planning model can coordinate these actions without requiring every detail to be pre-programmed. Instead, the robots infer the next steps from the state of their environment and the natural-language instructions they receive.

Apollo 2 was unveiled earlier this month alongside Apptronik’s new 90,000-square-foot Robot Park, where robots perform industrial tasks and generate real-world data used to train AI models. This close integration between hardware development and AI training is becoming common in the robotics industry. Rather than relying solely on simulated data, companies want their robots to practice in physical settings where they can encounter the unpredictability of real objects and human workspaces. The data collected from these trials can then be used to refine the AI models that control future generations of robots.

The robots still make mistakes

Despite the expanded capabilities, Google’s results show that the technology is not yet consistently reliable. Gemini Robotics 2 achieved a 92% success rate when unscrewing a lightbulb but only 36% when screwing one in. It succeeded 44% of the time when tying a trash bag and 32% when using a dustpan. These numbers illustrate an important distinction between impressive research demonstrations and production-ready systems. The model can perform many skills, but it cannot yet execute all of them with the consistency required for unattended operation. The variation in success rates also shows that certain tasks, particularly those requiring fine motor control with flexible objects, remain harder for current AI systems than others.

Google has also introduced ASIMOV-Agentic, a safety benchmark that tests whether robotics models can reject dangerous instructions, recognize when a task cannot be completed safely, and seek human assistance when uncertain. The company said the system can detect when a person moves too close to a robot and bring the machine to a safe stop. These safety features are likely to be critical as robots begin to operate in shared spaces with people. A humanoid robot moving around a warehouse or factory floor must be able to pause or stop when a worker enters its path. It should also refuse commands that could cause harm, such as asking it to strike a person or damage equipment. ASIMOV-Agentic is intended to measure those behaviors in a standardized way, giving developers a clearer picture of the risks before deployment.

Competition shifts to robotic software

The launch comes as competition shifts from building more capable robot bodies to developing the software that controls them. At least 13 embodied AI foundation models were released in June 2026. This explosion of models reflects a broader trend in the industry. Companies such as Amazon, Tesla, NVIDIA, and numerous startups are investing heavily in humanoid robots, but the underlying hardware is becoming more commoditized. The differentiator is increasingly the software stack that enables a robot to perceive, reason, and act in real time. Google’s move into this space with Gemini Robotics 2 is part of a wider race to create a general-purpose brain for robots, one that can be transferred across different machines without starting from scratch.

What Gemini Robotics 2 means for the future of robots

Gemini Robotics ER 2 is now available through Google AI Studio and in private preview through the Gemini Enterprise Agent Platform. Gemini Robotics 2 and its on-device version are currently limited to selected early-access partners. The release shows how robotics competition is shifting toward AI that can operate different machines across a wider range of tasks and environments. Google’s results also show that progress remains uneven, meaning humanoid robots may be getting smarter quickly, but they are not yet ready to work without testing, safeguards, and human oversight.

The broader implications go beyond factory automation. If Gemini Robotics 2 can generalize across robot platforms and tasks, the same underlying model might one day operate walking robots in logistics centers, assistive robots in homes, or inspection robots in hazardous sites. Google’s focus on multi-step reasoning and safety means that the AI is being designed not as a simple remote control, but as a cognitive layer that can plan, adapt, and explain its actions. However, the current reliability numbers suggest that more work is needed before such systems can be trusted with high-stakes decisions. The road from research demonstration to everyday deployment is long, but the arrival of full-body control marks a significant milestone in that journey.

Source: eWeek News