Gemini Robotics 2.5: Why Industry Still Can't Trust AI Yet

The 4% Gap in Industrial Automation
Google DeepMind’s latest update, Gemini Robotics 2.5, boasts significant gains in physical intelligence. By utilizing Gemini 4 Argon as a synthetic simulation engine, the model achieved a 96% success rate in automotive kitting tasks at Mercedes-Benz factories. While these numbers are impressive for general-purpose robotics, they highlight a critical reality for operations managers: the final 4% gap is a massive hurdle.
In a production environment, a 96% success rate is insufficient. As noted by Kanishka Rao, DeepMind’s robotics director, closing that remaining margin is a “steep road.” Unlike specialized, hard-coded industrial arms that operate with near-perfect consistency, Gemini-powered robots prioritize generalization—the ability to handle varied, unpredictable tasks—over the absolute precision required for high-volume assembly lines.
The Latency Problem
A major bottleneck remains the reliance on cloud infrastructure. Even with the performance improvements in the 2.5 release, the model does not run locally on the robot. Every action requires a round-trip to the cloud, introducing latency that makes real-time, high-speed interaction difficult. As Ars Technica pointed out, the promotional videos demonstrating these robots are often sped up by 2x to mask the processing pauses inherent in current agentic models.
How it works in practice
The system splits the workload into two distinct layers:
- Gemini Embodied Reasoning: A vision-language model that processes live video to plan the sequence of actions.
- Gemini Robotics 2.5 (VLA): A vision-language-action model that translates those plans into specific motor commands for actuators.
To improve safety, the model now includes inference-time reasoning. If the robot detects an obstacle or a human in its path, it will pause operations until the environment is clear. This feature reportedly increased safe task completion by 33% around obstacles and 43% near people.
Verdict: Who should use it?
Gemini Robotics 2.5 is currently a research-grade tool for partners like Apptronik and Boston Dynamics. It is not ready to replace human technicians in high-stakes manufacturing. If your operation relies on repetitive, high-speed tasks where a 4% failure rate would result in costly downtime, this technology is not yet a viable replacement. It is, however, a powerful platform for businesses looking to automate unstructured, low-stakes prep work that previously required constant human supervision.
Sources
- Gemini Robotics 2.5 accelerates real-world general dexterity — deepmind.google
- DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak" — arstechnica.com
Frequently asked questions
- Does Gemini Robotics 2.5 run locally?
- No. The model currently runs on cloud-based servers, which introduces latency that can affect real-time performance.
- Is Gemini Robotics 2.5 ready for factory use?
- It is currently being tested by partners in specific scenarios, but its 96% success rate is generally considered too low for critical, high-speed industrial production lines.
AI Briefing in 3 Minutes
Get what actually matters in AI & tech delivered free to your inbox. Zero spam, 1-click unsubscribe.
Comments
0 commentsDeixe seu comentário
Be the first to comment.
Continue Reading

Apple Starts Charging $50 for Extra AI Usage: What You Need to Know
Apple launches 50 dollar AI usage packs

Are Claude Code Settings Secretly Ditching Your Passwords on GitHub?
Claude Code leaks secrets to public GitHub

Why 74% of AI Coding Tasks Still Need Human Cleanup
New benchmark exposes limits of coding agents