Humanoid robots are nowhere near ready for real-world work, despite impressive demo videos circulating online. While artificial intelligence has made dramatic progress in knowledge work, physical robotics remains largely confined to controlled test facilities and carefully edited videos. The gap between what robots can do in labs and what they can do in actual homes, kitchens, and warehouses is vast, and the technical barriers are far more complex than most people realize.

What's Actually Stopping Humanoid Robots From Working?

The challenge isn't just building a robot that looks human. It's building one that can think, move, and adapt like a human. Steve Newman, who has analyzed the state of robotics development, identified 14 distinct technical hurdles that must be overcome before humanoid robots can function as general-purpose workers.

The most obvious barrier is the robot hand itself. The human hand is an engineering marvel, with roughly two dozen distinct joints and approximately 17,000 tactile sensors packed into a relatively small space. Current robot manipulators fall far short of this standard. While some experimental robot hands have achieved as many as 27 degrees of freedom, matching the human hand's overall combination of flexibility, sensitivity, strength, and reliability remains elusive.

Even more challenging is the control problem. A robot must not only have the right physical hardware; it must also solve incredibly complex computational problems in real time. It needs to figure out how to grasp a complicated object, plan a sequence of motions to fold a shirt or flip an omelette, and handle unpredictable materials like fabric or food that can shift suddenly.

How to Evaluate Robot Progress Beyond the Hype?

• Watch for Cherry-Picked Scenarios: Demo videos often show robots performing one successful task after many failed attempts, with camera cuts hiding struggles and slowdowns that would be visible in real time.
• Check for Environmental Manipulation: Impressive demos frequently arrange the scenario to avoid challenges the robot isn't ready for, such as cluttered spaces, unexpected obstacles, or materials the robot hasn't been trained on.
• Assess Speed and Reliability: Current humanoid robots often move much more slowly than humans, sometimes taking nearly two minutes to complete a single simple task like picking up clothing and placing it in a washing machine.
• Look for Real-World Deployment: The most telling sign of progress is whether robots are actually working in homes, factories, or hospitals, not whether they can perform isolated tasks in controlled environments.

Computer vision, which has advanced dramatically over the past decade, still struggles with real-world complexity. Picking out a specific object from a crowded environment, understanding where it should be grasped, and determining safe foot placement while avoiding knocking things over remains an unsolved problem at the level of reliability required for autonomous work.

Why Context and Common Sense Are Harder Than They Look?

Beyond physical manipulation, robots need to understand context and exercise judgment. A general-purpose robot must know where supplies are kept in a particular home, how a specific person likes their meals prepared, or whether a slight change in someone's gait signals a potential fall. Large language models like ChatGPT and Claude have made impressive strides in reasoning and planning, but the physical domain introduces additional layers of complexity.

One critical difference between AI systems and robots is the availability of training data. Large language models benefited from massive pools of pre-existing text, including a substantial fraction of all books ever written and the entire web. No equivalent dataset exists for physical tasks. There is no straightforward way to gather and label millions of examples of robots folding laundry, cooking meals, or repairing machinery in diverse real-world conditions.

Robots also struggle with speed and strength. While some industrial robots are extremely powerful, humanoid robots designed for general-purpose work typically aren't. Combining strength with manageable weight, a large number of joints, and a maneuverable frame is extraordinarily difficult. Powerful motors generate more heat and drain batteries faster, creating a fundamental engineering trade-off that hasn't been solved.

Coordination and adaptation present another barrier. Most AI agents operate in isolation within static environments. Robots, by contrast, will often need to cooperate with people or with other robots, and they'll need to re-plan constantly when things don't go as expected. A stuck bolt, a piece of rotten produce, or a child running into the kitchen all require the robot to abandon its original plan and adapt in real time.

The gap between impressive demo videos and functional artificial workers remains enormous. While the San Francisco AI scene widely believes that humanoid robots will soon enter everyday life, the physical side of AI remains mostly confined to test facilities. There is no robot equivalent to ChatGPT, nothing that ordinary people can access to see what these systems can actually do. Until that changes, progress will remain difficult to assess, and the timeline for truly capable humanoid workers will remain uncertain.