One of the most helpful pieces of design advice I heard from a senior engineer, early in my career, was: “You’ve planned for the best day. Now go plan for the worst.”
Today, as I watch teams build groundbreaking physical AI—systems designed to perceive, reason, and act in the physical world rather than exist solely in software, with applications ranging from industrial automation to self-driving cars to surgical robots—that lesson is more relevant than ever. More than $6 billion in capital flowed into robotics companies in the first seven months of 2025, and the investment pace has only accelerated. But many teams building robotics systems are designing for the best-case scenario, rather than planning for the worst. When these robots fail, the consequences are breakdowns, safety risks, and systems that people and organizations have come to depend on suddenly going dark.
According to the RAND Corporation, more than 80% of AI projects fail to reach meaningful production deployment, twice the failure rate of comparable non-AI technology projects. These are largely software deployments. The stakes are considerably higher when the system in question isn’t returning a bad search result or serving the wrong ad, but moving physical objects in a production environment, making split-second decisions in a surgical suite, or navigating a vehicle at highway speed .
Production turns “small misses” into significant costs
In the digital world, an AI error is a recoverable event. A model produces a bad recommendation; engineers retrain and redeploy. Users experience a brief degradation in quality. The feedback loop is fast, and the correction is invisible to most of the organization.
Physical AI operates under different constraints. When a robot misinterprets a torque specification, or a vision system fails to correctly identify a part under shifted lighting conditions, the consequences are immediate and material. The Siemens True Cost of Downtime 2024 report found that the world’s 500 largest companies lose approximately $1.4 trillion annually to unplanned downtime, representing 11% of total revenues. In automotive manufacturing specifically, an idle production line costs up to $2.3 million per hour.
Against that backdrop, the transition from lab-ready to floor-ready becomes something more than an engineering milestone. It is where business models are validated or broken. The standard the industry uses is the “three nines”—99.9% uptime or higher. That figure is often cited as a benchmark; what is less often discussed is what lies on the other side of it. At a facility running continuous shifts, the 0.1% that falls below three nines still represents nearly nine hours of potential downtime per year. At automotive-scale throughput, that is not an abstraction. It is a quantifiable cost with a quantifiable owner.
The factory floor is the proving ground
Research environments for robotics and physical AI are, by design, controlled. Lighting is consistent. Tools are maintained. Material lots are selected for their representative properties. The system is evaluated against distributions of data that closely resemble what it was trained on. This produces accurate performance measurements within those environments, and systematic blind spots for everything outside them.
Oliver Hsu, a partner at Andreessen Horowitz, who invests in emerging platforms and systems, described this dynamic in a 2026 analysis. A vision-and-language-action system that achieves 95% accuracy in a lab setting may degrade to 60 % accuracy when deployed, simply due to distribution shift—the gap between training conditions and real-world conditions. Even at 95% reliability in a high-throughput environment, Hsu noted, failures occur hundreds of times per day, requiring constant human intervention and effectively negating the productivity gains the system was deployed to create.
Factory environments introduce variables that compound this problem: lighting that shifts across shifts and seasons, tooling that degrades incrementally across production cycles, material lots with properties that vary within specification but outside the range the model has learned to handle. These are the normal conditions of industrial operation.
Design for the exception
The teams making the most durable progress in physical AI ask early in the design process: What does the system do when the model is uncertain?
That question sounds straightforward but many engineering teams encounter it for the first time after something has failed, rather than before they start designing.
Answering it upfront changes the architecture of the system in material ways. It requires defining confidence thresholds, not just accuracy targets. It requires building override logic and escalation paths before they are needed, not in response to an incident. It requires thinking about the interface between automated decision-making and human judgment as a designed handoff, not an emergency brake.
When confidence thresholds and human escalation protocols are built into the system from the outset, the result is a foundation for operator trust, which is the variable that most determines whether a physical AI deployment survives its first six months in production.
The competitive advantage in resilience
The industry’s attention is currently concentrated on model capability. That focus is producing genuine advances and obscuring a gap between what systems can do in demos and what they can sustain in production.
Bessemer Venture Partners described this gap in its 2026 robotics analysis: “We’re in the GPT-2.5 moment for robotics. Capabilities are real, but the gap between lab performance and field deployment remains wide.”
The companies that are closing that gap treat reliability as a designed property of the system, not an outcome of deploying a capable model. They treat human oversight as a structural feature, not a temporary accommodation. And they treat resilience—the ability to maintain acceptable performance when conditions degrade—as the measure by which a physical AI deployment will ultimately be judged.
The World Economic Forum’s analysis of physical AI in manufacturing found that digital twin simulations and AI-powered robots have delivered measurable gains, cutting deployment times by 40%, improving cycle times by 20-30%, and reducing error rates by 25%. But it cautioned that realizing those gains at scale requires embedding robotics in long-term operational strategy, not just pursuing short-term improvements. That distinction between a system that performs well when conditions cooperate and one that performs acceptably when they don’t, will define which physical AI deployments prove durable over the next decade.
The engineers who will build that future are the ones who have internalized the same lesson I learned early. The design work that matters most is the work you do for the worst day, not the best one.
Dave Evans is president and CEO of MISUMI Americas.