In early 2024, a robotics company running a pilot programme at a logistics facility disclosed that one of its humanoid units had collided with a worker during an unstructured handoff task, causing a minor injury. The company did not name the facility. It did not specify which robot was involved. And it did not say who — the manufacturer, the operator, or someone else — bore responsibility for the incident under the terms of the pilot agreement.

That disclosure, buried in a funding announcement, is about as much public information as exists on workplace incidents involving humanoid robots. There have been no regulatory filings, no published liability determinations, no court cases. The incidents are almost certainly happening — any technology being tested under real conditions will produce failures — but the legal and regulatory infrastructure to process them is only beginning to take shape.

As humanoid deployments move from narrow pilots to something more like ongoing commercial operations, the liability question is no longer theoretical. It is practical, urgent, and genuinely unresolved.

The Basic Legal Problem

Existing product liability law in most jurisdictions was built around a reasonably clear model: a manufacturer makes a product, a defect in that product causes harm, and liability flows from manufacturer to injured party. Courts have refined this framework over decades — distinguishing design defects from manufacturing defects, establishing standards for adequate warnings, and defining what "defective" means for complex machinery.

Humanoid robots strain this framework in several ways simultaneously.

First, these systems learn. A humanoid robot operating in a workplace today is not running purely on the software it shipped with. Its behaviour has been shaped by training data, refined through real-world operation, and may have been updated remotely since installation. When a robot trained on one set of conditions causes harm in a different set of conditions, the question of what constitutes a "defect" — and who introduced it — becomes genuinely complicated.

Second, the chain of deployment is long. A typical commercial humanoid involves a hardware manufacturer, one or more software providers (sometimes distinct from the hardware company), a systems integrator who customises the robot for a specific workplace, and the end operator who runs it. An incident might trace to a hardware failure, a software decision, an integration error, a training data problem, or an operator decision to deploy the robot in conditions outside its intended parameters. Attribution across that chain is not straightforward.

Third, the role of human oversight complicates everything. Most current humanoid deployments involve some degree of human supervision — an operator who can intervene, a remote monitor who can stop the robot if something goes wrong. If a human had the ability to prevent an incident and did not, how does that interact with the manufacturer's liability? Courts have not answered that question for humanoid systems specifically, because the cases have not come before them yet.

The Workers' Compensation Layer

In the United States, the most immediate practical framework for workplace robot incidents is workers' compensation — a system designed to handle industrial accidents regardless of fault, in exchange for limiting an employer's exposure to tort liability. If a humanoid robot injures a worker on the job, the first recourse is typically a workers' comp claim against the employer, not a product liability suit against the robot maker.

This works reasonably well for routine incidents. The worker gets compensated, the employer's insurer absorbs the cost, and rates adjust over time to reflect the risk profile of the workplace. The manufacturer is largely insulated from direct liability unless the employer or insurer chooses to bring a separate product liability action — which requires proving a defect, establishing causation, and clearing other legal hurdles that make it a more expensive and uncertain path.

The consequence of this structure is that most humanoid workplace incidents will likely be handled through workers' comp, quietly, without generating public records or legal precedents. That is good news for manufacturers and operators in the short term. It is less good for the development of a coherent legal framework, because the incidents that would test and clarify the law are being absorbed into a system that discourages the litigation that produces precedent.

Workers' comp also has limits. It covers physical injuries, not all harms. It does not address scenarios where a robot's actions cause property damage to a third party, or where a worker is injured by a robot owned by a contractor rather than their direct employer. The more complex the deployment environment, the more likely it is that real incidents fall through the gaps in the existing framework.

What the EU Is Doing — and What It Leaves Out

The European Union's AI Act, which came into force in 2024 and is being phased in through 2026 and 2027, is the most substantive regulatory framework for AI-enabled systems that currently exists. It classifies AI applications by risk level and imposes requirements accordingly — transparency obligations, human oversight requirements, and prohibitions on specific high-risk uses.

Humanoid robots operating in workplaces fall into a category the AI Act treats as high-risk, which means manufacturers must meet documentation, testing, and monitoring requirements before deploying them commercially in EU member states. That is a meaningful threshold. It means the legal risk of deploying an inadequately tested system is real and enforceable, not merely theoretical.

What the AI Act does not do is resolve liability when something goes wrong. It establishes standards and imposes obligations on manufacturers and operators. It does not create a direct right of action for injured individuals, nor does it clearly specify how liability should be apportioned across the deployment chain when an incident occurs. Separate EU proposals on AI liability — which would make it easier for injured parties to access information and bring claims — have been in legislative process for several years without producing final rules.

The result is a system that has begun to regulate humanoid deployment without fully addressing what happens when deployment causes harm. That is not unusual for emerging technology regulation; the rules tend to lag the technology. But it means that even in the jurisdiction with the most developed AI regulatory framework, the liability question remains open.

The Insurance Gap

Behind the legal framework sits a practical question: who is actually underwriting the risk of humanoid deployment, and at what price?

Commercial general liability insurance — the standard coverage that businesses carry for third-party bodily injury and property damage — will cover some humanoid robot incidents under existing policies. But insurers are actively working out how to price and scope humanoid-specific risk, and the answer varies considerably by carrier and by the nature of the deployment.

Some insurers are excluding AI-driven robotics from standard policies and requiring separate riders. Others are offering coverage but with significant uncertainty about how claims would be handled, given that the underwriting models were built on historical data that predates autonomous humanoid systems. A few specialist insurers are beginning to offer product-specific coverage for robotics companies, but the market is thin and premiums reflect the uncertainty.

For operators deploying humanoids commercially, this creates a practical problem. The risk profile of these systems is not well-characterised enough for insurers to price it accurately. Operators may find themselves either paying for coverage that is vague about what it actually covers, or accepting gaps in coverage that they do not fully understand. Neither is a stable situation as deployment scales up.

What Manufacturers Are Actually Doing

In the absence of settled law, humanoid manufacturers have been managing liability exposure through contract. The agreements between manufacturers and commercial operators typically include indemnification clauses, limitations on use cases, requirements for human supervision ratios, and restrictions on operating conditions. A robot sold for warehouse use in a controlled environment will typically have contractual terms that exclude liability if the operator deploys it on a construction site or in a public space.

This is sensible risk management, but it has a structural weakness: the contracts define risk allocation between the parties to the agreement, not liability to third parties who were not at the table. A warehouse worker injured by a humanoid robot has no contract with the robot's manufacturer. Their recourse flows through workers' comp, through their employer, and potentially through tort law — none of which neatly maps onto the contractual risk allocation that the manufacturer and operator negotiated.

Manufacturers are also, quietly, building data logging and incident reporting infrastructure into their systems. Every significant humanoid platform now generates logs of robot behaviour, sensor readings, and decision points that can, in principle, be used to reconstruct what happened in an incident. Whether those logs will be discoverable in litigation, and how they will be interpreted, is an open question — but their existence is a sign that manufacturers are preparing for the legal scrutiny that will eventually come.

The Autonomous Decision Problem

All of the above gets more complicated as the autonomy level of these systems increases. Current commercial humanoids are, in meaningful ways, supervised: they operate in constrained environments, on defined tasks, with humans available to intervene. As the technology develops, the expectation is that the degree of human oversight will decrease — that robots will handle more edge cases autonomously, operate in less structured environments, and require intervention less frequently.

That trajectory creates a liability problem that existing frameworks are not equipped for. When a robot makes an autonomous decision that leads to harm — not because of a hardware defect or a software bug in the traditional sense, but because its learned decision-making process produced an action that a reasonable person would not have taken — who is responsible? The manufacturer who trained the model? The operator who deployed it in that environment? The developer who designed the decision-making architecture?

Legal scholars have been writing about this problem for years, typically under the heading of the "responsibility gap" in AI systems. The concern is that sufficiently autonomous systems may produce outcomes where no one is clearly to blame under existing legal concepts, even when harm has clearly occurred. The practical response to date has been to keep humans in the loop specifically to preserve clear liability chains — a strategy that works until the economic pressure to reduce human oversight becomes strong enough to override the legal caution.

The liability question will not be resolved before humanoid deployment expands. It will be resolved — partially, messily, through litigation and legislative response — in the wake of incidents that are now becoming increasingly likely. The legal and regulatory uncertainty is not an obstacle to deployment so much as a condition of it. The companies, regulators, and insurers that engage seriously with this question now will be better positioned when the first significant cases arrive. The ones that treat it as someone else's problem are accumulating risk they have not yet fully priced.