A new warehouse worker on their first shift is slow, cautious, and makes mistakes. By the end of the first week, they're faster. By the end of the first month, they've developed a feel for the quirks of their specific facility — which shelving units are slightly misaligned, which conveyor belt has a predictable hiccup, which colleagues to check with when something doesn't look right. They are, in a meaningful sense, a different and more capable worker than the person who showed up on day one.
A humanoid robot deployed in that same warehouse does not generally work this way. In most current deployments, the robot that finishes its hundredth shift is essentially identical — in terms of capability — to the one that started its first. The experience hasn't accumulated. The small corrections haven't been retained. Nothing about the specific environment has been internalized in a way that makes the robot better at operating within it.
This is the memory problem in humanoid robotics, and it receives far less attention than questions about locomotion, grasping, or battery life. But it is arguably as important as any of them — and in some respects harder to solve.
What "Learning" Actually Means in This Context
It's worth being precise about terminology, because the word "learning" is used loosely in robotics coverage in ways that obscure more than they reveal.
Most current humanoid robots are trained — extensively — before deployment. That training process involves exposure to large amounts of data: simulation environments, teleoperation demonstrations performed by human operators, video footage of tasks being completed, and increasingly, large-scale neural network models (often called foundation models or vision-language-action models) that encode general knowledge about the physical world. By the time a robot arrives on a factory floor, it has processed an enormous quantity of information about how objects behave, how bodies move, and how tasks are structured.
What the robot typically does not do is continue learning in a meaningful way once that deployment begins. The weights — the numerical parameters inside the neural networks that encode everything the robot "knows" — are frozen. The robot is executing a fixed policy, not refining one. When it encounters a situation its training didn't adequately prepare it for, it fails, and that failure does not automatically improve its performance the next time the same situation arises.
This is a deliberate design choice as much as a technical limitation. Updating a deployed robot's parameters based on real-world experience raises serious safety and reliability questions. A robot that is continuously modifying its own behaviour in response to environmental feedback is a robot whose future behaviour becomes harder to predict and audit. In regulated industries, that predictability is not optional.
The Catastrophic Forgetting Problem
When researchers attempt to enable robots — or AI systems more broadly — to learn new tasks after initial training, they run into a well-documented phenomenon called catastrophic forgetting. The neural network, in learning to handle a new situation, overwrites the parameters it previously used to handle old ones. The robot becomes better at the new task and worse at everything it previously knew how to do.
Catastrophic forgetting has been studied in machine learning since the late 1980s, and it remains unsolved in any general sense. Partial mitigations exist: techniques like elastic weight consolidation, which penalises large changes to parameters important for previous tasks, or progressive neural networks, which add new network components for new tasks without modifying existing ones. But these approaches introduce their own complications — increased computational overhead, architectural constraints, and in practice, imperfect retention of old skills even under the best conditions.
For humanoid robotics, this isn't an abstract research concern. It's a practical obstacle to the kind of incremental on-the-job improvement that makes human workers valuable over time. A robot that gets better at Task A by forgetting how to do Task B is not a robot that's learning in any useful operational sense.
The Fleet Learning Promise and Its Limits
Several humanoid robotics companies have proposed fleet learning as a path through this problem. The concept is appealing: rather than each robot learning only from its own experience, all robots in a deployment share what they've encountered. A fleet of a thousand units, collectively logging millions of hours of operation, generates a dataset far richer than any single unit could accumulate. That data can be used to retrain the shared model, which is then updated across the fleet. Each robot benefits from what every other robot has experienced.
This is a coherent approach, and for narrow, well-defined tasks it has some operational precedent in non-humanoid industrial robotics. But it has several constraints worth understanding clearly.
First, fleet learning as described above is not continuous on-device learning — it's centralised periodic retraining. The individual robot is still running a frozen policy during its shifts. The improvements happen between training cycles, not within them. The robot is not adapting in real time to its environment; it's receiving periodic updates based on what the broader fleet encountered.
Second, fleet learning depends on fleet size. Companies like Physical Intelligence, Figure AI, and 1X Technologies have all gestured toward this approach, but none has publicly demonstrated it at meaningful scale. The hypothesis that fleet data will produce robust, generalising improvements remains largely unvalidated outside research settings.
Third, there's a data quality problem. A robot fleet logging failures is not the same as a fleet generating useful training signal. Knowing that a robot dropped a box 847 times under various conditions is only useful if those failures are labelled, contextualised, and representative of the distribution of real-world challenges. Raw operational data from deployed robots is messy in ways that curated training data is not.
Simulation as a Partial Answer
One approach that has gained significant traction is the use of simulation to extend effective training before and between deployments. If real-world experience is expensive to collect and difficult to learn from continuously, simulated experience is cheap and controllable.
Modern physics simulators — tools like Isaac Sim from Nvidia, or MuJoCo, which has long been a standard in robotics research — can generate enormous quantities of synthetic training data at speeds impossible in the real world. A robot that would take years to accumulate a million hours of physical experience can log that time in simulation in days.
The limitation is what researchers call the sim-to-real gap: behaviours learned in simulation don't always transfer cleanly to the physical world, because no simulator perfectly replicates the texture, friction, deformation, and unpredictability of real materials and real environments. The gap has narrowed substantially over the past several years, particularly as simulators have improved their physical fidelity and as domain randomisation techniques — deliberately introducing variation into simulated environments to make trained behaviours more robust — have matured. But it hasn't closed. The gap shows up in exactly the situations that matter most: edge cases, unusual configurations, and the kinds of low-probability events that cause deployment failures.
What Would Actually Solve This
The research directions that seem most promising don't involve a single breakthrough so much as an accumulation of partial solutions. Continual learning research is making incremental progress on catastrophic forgetting, though no approach has yet produced systems reliable enough for wide deployment. Memory-augmented neural networks — architectures that separate a trainable model from an external, updateable memory store — offer one way to add adaptability without destabilising the underlying policy. And the continued improvement of foundation models means that robots may arrive at deployment with sufficiently broad prior knowledge that on-the-job adaptation matters less than it otherwise would.
That last possibility is worth sitting with. If a model trained on an enormous diversity of physical environments generalises well enough, the need for environment-specific adaptation may be lower than the current generation of narrow, task-specific systems would suggest. Several labs are betting on this. Whether the bet pays out at the level of reliability that commercial deployment requires is genuinely unknown.
What's clear is that the question of whether a humanoid robot can get meaningfully better at its job over the course of a deployment — retaining what it's learned, building on it, and improving in the ways that make human workers increasingly valuable over time — remains open. It's a less photogenic problem than the ones that generate press coverage. But for anyone evaluating whether humanoid robotics will deliver on its operational promises, it's one of the right questions to be asking.