AI has advanced rapidly in work that happens inside a computer. Robotics is pursuing a more difficult target: systems that can reliably perform useful physical work in changing, unstructured environments.
That distinction matters for executives and builders evaluating humanoids and other general-purpose robots. A striking video can demonstrate a real capability, but it may not show the repetition rate, setup required, recovery from errors, operating speed or safety margins needed for an economic deployment.
The demo-to-deployment gap
Robotics progress is often communicated through curated demonstrations. Those clips can be valuable, but they make it difficult to assess the variables that determine commercial value: how often the task succeeds, what happens when objects are misplaced, whether the environment was arranged for the machine, and how quickly it can recover from a failure.
For buyers, the relevant question is not simply whether a robot can complete a task once. It is whether it can complete a defined workflow repeatedly, at acceptable throughput and cost, without requiring more human supervision than it displaces.
Public competitions and field trials may offer a clearer signal than polished videos because they reduce the opportunity for cherry-picking. Even then, a single performance is not the same as evidence of durable operation across shifts, sites and edge cases.
Hands, vision and control remain coupled problems
Human hands combine many joints, fine motor control and extensive tactile sensing. Robotic hands can match parts of that profile—such as high numbers of degrees of freedom or dense tactile sensors—but combining dexterity, sensitivity, strength, durability and manufacturability remains difficult.
The harder problem is not just building the hand. A robot must use it well: identify a safe grasp in a cluttered scene, control force on fragile or deformable materials, and coordinate many joints while reacting to unexpected movement. Folding clothing, handling food, fastening parts in tight spaces and manipulating flexible materials are all examples where small errors can compound.
Vision introduces another layer. Modern computer vision is powerful, but a workplace robot must reliably distinguish objects in clutter, estimate their position, understand hazards and avoid knocking over nearby items. A warehouse aisle, a commercial kitchen and a care facility provide far less consistency than a controlled test setup.
Intelligence needs physical context
General-purpose work also requires task planning and replanning. A robot needs to decide how to access a tool, clear a path, sequence actions and respond when a bolt is stuck or an item is missing.
Language models may help with reasoning and instructions, but physical learning has a data disadvantage. Models for knowledge work can draw on enormous pools of text and code. There is no equivalent corpus for the full range of physical interactions, object properties and workplace exceptions. That makes efficient learning, generalization and adaptation on the job central challenges.
Robots will also need operational context: where supplies are stored, site-specific procedures, user preferences and signals that an unusual situation needs human intervention. In many settings, they must coordinate safely with people and other machines rather than operate alone in a static environment.
Economics are constrained by speed, strength and energy
A robot that moves slowly may be suitable for overnight cleaning or other low-urgency tasks. But speed affects economics. A machine that takes too long to handle an item will struggle to justify its capital cost in a warehouse, kitchen or care role—and may become an obstacle to human workers.
Increasing speed is not a simple software upgrade. Faster movement leaves less time to perceive and react, raises control demands and increases the consequence of collisions. Strength creates similar trade-offs: larger motors add heat, consume more battery power and can make a mobile robot heavier and harder to operate safely.
Mobility is another unresolved design choice. Wheeled systems tend to be more stable, cheaper and able to carry more weight. Legged systems can traverse stairs, clutter and uneven environments, but add complexity and fall risk. The optimal form factor will likely remain application-specific.
What to watch next
The near-term opportunity is likely to favor constrained workflows over all-purpose humanoid labor. Buyers should ask vendors for evidence on task success rates, human intervention, recovery behavior, uptime, throughput, safety incidents and the environmental changes the system can tolerate.
The most consequential milestone will not be another isolated feat of dexterity. It will be sustained, measurable performance in real facilities—where reliability, integration and unit economics matter as much as the robot’s ability to move.
