The robot industry is no longer chasing the fastest or most dexterous machines; instead, companies are racing to build systems that work reliably in the real world, stay safe, and actually make money. This shift became clear in late July 2026, when Google DeepMind unveiled Gemini Robotics 2 alongside a wave of commercial robot deployments across Japan and Europe, signaling that physical artificial intelligence (AI) has entered a new competitive phase focused on integration, safety assurance, and operational sustainability rather than isolated performance breakthroughs.

What Changed in the Robot Foundation Model Race?

On July 30, Google DeepMind announced Gemini Robotics 2, a suite of AI models designed to handle the messy reality of robot work. Unlike earlier systems optimized for single tasks, Gemini Robotics 2 divides intelligence across three layers: a whole-body vision-language-action (VLA) model that controls robots like the Apptronik Apollo 2 for complex manipulation involving walking and bending; a higher-level Executive Reasoning (ER) model that orchestrates multi-stage tasks lasting several minutes; and an on-device adaptation system that learns new hardware configurations with fewer than 200 examples and just a few hours of data.

The performance metrics are striking. The system achieved an 89.6% success rate for precision insertion tasks with a Franka Duo robotic arm and a 92% success rate for lightbulb removal with the Apollo 2 humanoid. More importantly, Gemini Robotics 2 can coordinate multiple robots working together and adapt to new hardware without extensive retraining, capabilities that matter far more in factories and warehouses than raw speed alone.

Why Is Safety Becoming the New Competitive Battleground?

Alongside the performance announcements, Google DeepMind released the ASIMOV-Agentic benchmark, a framework for testing robot safety. This signals a critical realization: as robots move from labs into hospitals, warehouses, and homes, the industry cannot afford security failures. A research team published a comprehensive survey on arXiv documenting how attacks on world models (the AI systems that help robots understand their environment) can propagate through an entire robot's decision-making pipeline and result in dangerous physical actions.

The survey identified multiple attack vectors that threaten embodied AI systems throughout their lifecycle. These vulnerabilities include data poisoning, backdoors in training data, adversarial inputs designed to fool perception systems, sensor spoofing, prompt injection attacks, trajectory tampering, and supply chain compromises. Each of these can corrupt the robot's internal understanding of the world, leading to incorrect predictions about how actions will affect the physical environment.

Researchers also flagged a subtle but dangerous risk: world models can create a "predictive safety illusion" where robots over-trust their own incorrect predictions, leading to confident but wrong decisions. To counter these threats, the research team proposed defensive measures including provenance management to track data origins, robust grounding to verify sensor inputs, uncertainty evaluation to flag low-confidence predictions, trajectory gates to block dangerous actions, and feedback auditing to catch errors in real time.

How Are Companies Monetizing Physical AI Beyond Prototypes?

The shift toward profitability became visible in three major commercial announcements in late July. NTT West, a Japanese telecommunications company, launched the "Omakase Inspection Robot Pack," a flat-rate service combining the autonomous mobile robot ugo mini with AI, cloud services, and maintenance support. The service costs 234,080 yen per month (roughly $1,600 USD) with a minimum three-year contract, targeting electrical rooms, data centers, and equipment rooms for automatic patrols, anomaly detection, and report generation.

What makes this significant is that it is not a robot sale; it is a managed service. The company handles installation, maintenance, education, and operational infrastructure, removing the burden from customers. This model mirrors how cloud computing and software-as-a-service (SaaS) businesses operate, suggesting that the most profitable robot companies may not be those selling hardware but those providing ongoing operational support.

KION Group, a major logistics automation company, reported an 18% revenue increase in its Intelligent Automation Solutions division during the first half of 2026, with adjusted earnings before interest and taxes (EBIT) jumping 35% year-over-year. The company acquired a 70% stake in Smart Innovation NV, a research subsidiary of Belgian retailer Colruyt, to accelerate development of autonomous pallet trucks capable of automatic loading and unloading. This acquisition signals that warehouse automation demand is no longer theoretical; it is translating into real sales and profits.

What Role Are Domestic Manufacturers Playing?

Japan and Europe are not waiting for US companies to dominate physical AI. KDDI and AVITA, a Japanese robotics firm, demonstrated a domestic humanoid robot equipped with Google's Gemini 3.1 Flash Live Preview conversational model at Google Cloud Next Tokyo. The key innovation was real-time voice processing that captures tone, speed, inflection, and emotional nuance without the multi-second delays that plagued earlier systems relying on cloud APIs.

RobotBank began offering an improved version of STARWALK, a transport robot that autonomously follows people to carry luggage, optimized for the Japanese market by switching to Japanese-made batteries. The focus on domestic supply chains and maintenance infrastructure reflects a broader recognition that robot adoption depends not just on performance but on local support ecosystems.

ELSA Japan acquired exclusive domestic sales rights for DEEPRobotics' Lynx S10, a wheeled quadruped robot significantly smaller than its predecessor. The S10 measures approximately 60 centimeters long, weighs 19 kilograms, and features 16 degrees of freedom, a top speed of 18 kilometers per hour, and a maximum payload of 8 kilograms. It can climb continuous 20-centimeter stairs and single 50-centimeter steps, operates in temperatures from minus 20 to 55 degrees Celsius, and runs for up to 3 hours or 16 kilometers on a single charge. Domestic delivery is scheduled for November 2026 or later.

How to Evaluate the Real-World Readiness of Physical AI Systems

• Operational Integration: Look beyond isolated performance benchmarks. Does the system include maintenance, cloud services, installation support, and ongoing operational infrastructure, or is it just hardware? Companies offering managed services are signaling confidence in reliability and long-term viability.
• Safety Assurance: Check whether the robot maker has published security frameworks, vulnerability assessments, or safety benchmarks. The presence of formal safety testing indicates the company understands real-world deployment risks and is taking steps to mitigate them.
• Domestic Supply Chains: Examine whether the manufacturer has localized battery sourcing, maintenance networks, and support teams. Robots that rely entirely on overseas supply chains face adoption barriers in regulated industries like healthcare and critical infrastructure.
• Multi-Robot Coordination: Assess whether the system can work alongside other robots or existing equipment. Single-robot deployments are limited; systems that coordinate multiple agents unlock warehouse and factory-scale applications.
• Adaptation Speed: Determine how quickly the robot can learn new tasks or hardware configurations. Systems requiring weeks of retraining are impractical; those adapting in hours or days with minimal data are production-ready.

The competitive axis of physical AI has fundamentally shifted. In 2024 and 2025, the headlines focused on which company could build the fastest humanoid or the most dexterous hand. By mid-2026, the conversation has moved to hierarchical intelligence, field operations, safety assurance, and business profitability. Google DeepMind's Gemini Robotics 2 exemplifies this shift by dividing labor among specialized models rather than chasing a single all-purpose system. Meanwhile, companies like NTT West and KION are proving that the real money lies not in selling robots but in operating them reliably, safely, and profitably over years.

For investors, entrepreneurs, and enterprises evaluating robot deployments, this transition carries a clear message: the winners in physical AI will not be those with the flashiest demos but those who can integrate safety, operations, and profitability into a coherent business model. The age of robot performance theater is ending. The age of robot operations is beginning.