Humanoid Robots Enter Factories Faster Than Safety Rules Can Follow

Agility Robotics and FORT Robotics announced a partnership in October to harden the safety architecture of Agility's Digit 5 humanoid robot — and behind that engineering deal sits a question the industry still cannot answer cleanly: who is liable when a mobile humanoid fails on a factory floor?

The partnership covers onboard safety systems, external interfaces, communications, and manual override. That scope itself signals a shift. Making robots safe in human workplaces now means far more than making the robot itself safe.

The stakes are straightforward. A robot can pass a performance test and still become a hazard the moment it meets an unexpected obstacle, receives a software update that changes its behavior, or encounters a real worker. As the source piece puts it, the question is no longer whether humanoid robots can perform a task, but whether they can do it safely — and who answers when they don't.

Why Humanoids Break the Old Safety Model

Traditional industrial robots live in enclosed work cells, separated from workers by barriers and safety devices. Humanoids do the opposite by design. They navigate environments built for humans, switch between tasks, and adapt to changing layouts.

That flexibility is exactly what creates the new risk surface:

• A false detection can trigger an unnecessary physical movement.
• A software update can alter how the robot behaves after validation.
• A changed factory layout can produce hazards that earlier tests never exercised.

Dana R. Wolfe, Vice President and Head of Industry at GFT Americas, told Unite AI that preventing unsafe interventions requires multiple layers of protection: deterministic software guardrails, factory-specific confidence thresholds, sensor fusion, and staged validation before deployment.

"There is no way to completely avoid false-positive results," Wolfe said. Instead, manufacturers have to ensure that mistakes don't translate into unsafe physical movements.

Her prescription is layered: confidence thresholds and human authority belong inside the system's safety architecture, alongside hardware interlocks. In plain words — the robot should doubt itself, and a person should always be able to stop it.

Do Current Standards Cover Mobile Humanoids?

Not fully, and that gap is the regulatory heart of this story. The International Organization for Standardization (ISO) has published safety requirements for industrial robots, and ISO 10218-2 covers integrating robots into complete systems. Those standards provide what Wolfe calls a valuable foundation — but they do not address all applications or hazards of mobile robots working alongside humans.

Wolfe argues that independent testing, scenario-based validation, and phased workplace pilots matter most while industry practices are still evolving. She also draws a line that many deployments will find uncomfortable: safety evaluations must be repeated whenever the AI model, the software, movement restrictions, the task, or the factory environment change in ways that could invalidate earlier results.

The mechanism here is simple. A robot that was safe under one set of conditions may not be safe after a major change — meaning safety becomes a lifecycle obligation, not a one-time certification. That is a structural cost of putting adaptive AI into physical work.

What Does Human Oversight Actually Have to Mean?

Monitoring every decision an AI system makes is impractical. But when humans are cut out of critical decisions, accountability becomes much harder to prove — a tension the industry has not resolved.

Adrian Figueroa Ramos, co-founder and CEO of Flexio and Routina, said oversight should concentrate on the decisions carrying the most uncertainty or risk.

"In today's AI world, the volume of outputs is more than a human can process. Reviewing everything makes the human the bottleneck, but human judgment is still necessary for high-risk or low-confidence decisions," he told Unite AI.

For physical AI, that principle gets concrete fast. Human oversight should mean named individuals with the power to override a system, and decisions that are documented and checked — not a supervisor clicking an approval button. If nobody with authority is identifiable at the moment of failure, liability dissolves into the workflow.

Who Pays When a Robot Fails?

The accountability framework proposed in the piece distributes responsibility according to how much control each party has over the system. Consider the scenario where a robot acts on a faulty component misidentified by its AI: the manufacturer, the software developer, the system integrator, and the employer using the system could all be held liable.

The proposed split of duties:

• Manufacturers create and document safety measures.
• Software developers test changes that could affect perception or movement.
• Integrators validate the whole system in its intended environment.
• Employers train workers, ensure safeguards exist, and maintain procedures for stopping unsafe operations.

Cybersecurity sits inside this picture too, not beside it. Operational technology connected to robots can be vulnerable through sensors, software, and external systems. A security failure that changes physical behavior is a workplace safety problem, not just an IT problem — which pulls security teams into the same accountability chain as robot engineers.

Why Performance Won't Decide Adoption

Humanoid robots are expected to become more flexible and more productive in factories. But the source piece's central argument is that adoption will not be decided by performance metrics alone. What's missing is clear responsibility, independent validation, and safeguards sustained across a robot's entire working life.

The closing framing is blunt: yes, a robot can make mistakes. The real question is whether the people using it have built a system that will hold someone accountable for those errors.

For factories weighing humanoid deployments now, that means the October Agility–FORT partnership is best read not as one vendor's engineering choice, but as a template: onboard safeguards, override channels, staged validation, and a written map of who answers for what. The companies that sort out accountability before the first incident will be the ones that get to scale.