DeepMind Reveals What Gemini Robotics 2.5 Can Do

Gemini Robotics 2.5 is unveiled. It's DeepMind's advanced AI platform, its latest iteration, bringing robotic capabilities strikingly close to human proficiency for real-world tasks, a major advancement in artificial intelligence. Gemini-powered robots are already performing practical duties within Mercedes-Benz factories. And DeepMind also introduced an open-source safety benchmark to promote a secure environment for future human-robot collaboration.

Advanced Dexterity and Speed Mark New Era

DeepMind's platform is complex. Gemini Embodied Reasoning is its brain, a vision-language model designed to interpret intricate instructions and formulate effective plans by processing continuous live video feeds from the robot's surroundings. This helps the robot grasp its environment. And working in concert is Gemini Robotics, the vision-language-action module we've upgraded to version 2.5, which directly generates the physical movements, actuating a robot's joints. It's very important.

Gemini Robotics 2.5 is better. The model demonstrated a notable improvement in performance, particularly for jobs requiring fine motor skills. During evaluations covering over 200 different tasks, it handled complex actions like serving food with tongs or precisely using a pipette, a common laboratory tool, without any errors. It's truly remarkable. A particularly impressive demonstration involved an Apptronik Apollo 2 humanoid robot, equipped with Gemini Robotics 2.5, successfully tying a bow using its five-fingered hands, an achievement that matters greatly because handling deformable objects such as ribbon has long been extremely difficult for robotic systems.

Tests are here. Lab results indicate the new version performs 30 to 100 percent better than the old one. This is a major advancement, pointing to a strong improvement in accuracy and speed for self-governing systems and expanding their capabilities. It's also 20 to 30 percent faster.

Real-World Performance and Remaining Hurdles

Labs weren't enough. Apollo 2 robots deployed. So, DeepMind partnered with Apptronik and Mercedes-Benz, intending to test its AI in real-world industrial settings, where these Gemini-powered machines were tasked with complex part kitting operations across several European automotive factories. They achieved 96 percent success. This marked a considerable advance compared to earlier robotic AI systems in similar scenarios, indicating a major step forward for factory automation. But, while that 96 percent is impressive, it still leaves a small but very important difference when measured against human efficiency on the same tasks.

Humans achieve near-perfect accuracy. Kanishka Rao, a director at DeepMind, acknowledged this remaining challenge, stating closing the final 4 percent gap will be a "steep road." Traditional industrial robots are task-specific. They're programmed for high accuracy, a form of automation that's been commonplace for decades. DeepMind's ambition with Gemini Robotics is different; it aims to create general-purpose robots capable of handling a vast array of activities, rather than just a few specialized functions. But this pursuit of generalization comes with its own trade-offs.

We can be very good at a particular application at the cost of generalization," explained DeepMind Robotics head Carolina Parada. Choose wisely. But we can also push on the generalization axis, she added, accepting that we might not get to 99.9 percent in every application if that's our focus.

Training Data and Latency Considerations

Essential training data drives all generative AI models. Text-based chatbots draw from the internet's vastness, but "physical data" for robotic AI, it's far scarcer. Teleoperation provides the best data. Humans guide robots. To accelerate Gemini Robotics 2.5's development, the team utilized Gemini 4 Argon, which created a "simulation engine" that generated thousands of practice run variations based on real-world camera data. It created "digital twins." These virtual replicas helped Gemini Robotics 2.5 generalize quickly to new tasks and environments. This synthetic data was then combined with traditional training demonstrations to ground the model in reality. And Argon wasn't just for generation; it also proved invaluable for analyzing failures and suggesting improvements in both simulated and physical robot operations.

It's not perfect. Frontier models, such as Gemini, need processing time for their tokens, which often introduces noticeable latency into their responses, hindering immediate interaction. The VLA within Gemini Robotics 2.5, while demonstrating improvement over earlier iterations, continues to rely on cloud-based servers, meaning it's not entirely integrated directly onto the robot itself, impacting real-time responsiveness. This dependency means reported improvements in both accuracy and speed are still subject to network latency, a critical concern for fluid real-time operation. The video you see is sped up. Two times. This makes the robot appear more human-like.

Prioritizing Safety with ASIMOV-Jailbreak

Ensuring the safety of robots operating near humans is paramount. DeepMind has incorporated inference-time reasoning into Gemini Robotics 2.5, enabling it to detect and avoid hazards. If an obstacle or person enters its operational space, the robot will either halt completely or wait until it is safe to resume its task. This safety feature significantly improves real-world deployment, with Gemini robots completing tasks safely 33 percent more often when obstacles were present, and a remarkable 43 percent increase when people were in the way.

DeepMind has initiated ASIMOV-Jailbreak. It's an open-source database. This initiative aims to standardize safety testing in physical AI, pushing the industry forward. It contains 1,500 prompts. These are engineered to trick robotic AI into performing unsafe actions. For instance, a prompt might instruct a robot to place items into a bag and then close it with a fastener too delicate for the robot to handle safely. But the desired outcome is for the AI to decline the task rather than attempting it and potentially causing damage. Integrating Argon into the loop yielded an impressive 98.8 percent jailbreak safety level, significantly higher than the 73.5 percent achieved by Gemini Robotics ER 2 alone in Google's safety benchmark. The full ASIMOV dataset is publicly available on Hugging Face for researchers and developers working on embodied AI.

• Gemini Robotics 2.5 achieves 30-100% higher success rates.
• It operates 20-30% faster than the previous version.
• Apollo 2 robots with Gemini Robotics 2.5 achieved a 96% success rate in Mercedes-Benz factories.
• Inference-time reasoning improved safe task completion by 33% with obstacles and 43% with people.
• ASIMOV-Jailbreak with Argon in the loop scored 98.8% in jailbreak safety.

While Gemini Robotics 2.5 is not yet generally available, it is actively being deployed with Google's robotics partners, including Agile Robots, Apptronik, and Boston Dynamics, setting the stage for its broader impact.