The five-level robotics autonomy framework: where humanoids actually stand today

Summary

Humanoid robotics discourse fixates on level-five autonomy, but the technology path is messier. Semi Analysis's five-level autonomy framework, borrowed from autonomous vehicle standards, clarifies where robotics actually stands and why most humanoid demos mislead.

Level zero is scripted motion. Factory robots perform the same programmed task millions of times. No cameras, no machine learning. A 50-ton arm requires light curtains and constant human oversight. Many facilities shut them down during lunch breaks despite theoretical 24/7 capability. Maintenance is so intensive that continuous operation is often uneconomical.

Level one is intelligent pick-and-place. A gantry with motors on X and Y axes, a camera, and basic computer vision using standard OpenCV and object detection. It's the carnival claw game scaled up. Companies demonstrating humanoid robots performing this task while claiming level-four or level-five autonomy are misleading. Using a humanoid for what a gantry arm does is economically wasteful and like constraining an LLM to text-only output just to prove it's powerful. The humanoid should call on a standard robotic arm when the task warrants it.

Level two is autonomous mobility. Scene understanding, open-world navigation, terrain traversal. Boston Dynamics' robot dog exemplifies this level. These machines are already deployed for narrow, high-value use cases: inspecting nuclear power plants, oil refineries, construction sites, critical infrastructure. A Chinese robotics firm demonstrated a bot that can walk like a spider, swim underwater, and take flight for tasks requiring adaptive movement in unstructured environments. The jobs are real and deployed, though limited in scope.

Level three is low-skill manipulation. Generalizable pick-and-place confined to a fixed zone. Mobile manipulation picks up a box and moves it across a space with general oversight. Bipedal humanoids, which don't need five fingers or human appearance, are already in pilot stages in kitchens, laundromats, manufacturing, and logistics. Remote operation is entirely valid here: one human monitoring multiple robots on screens and jumping between viewpoints as needed. A remotely-operated humanoid with dexterous hands would unlock genuine value, such as picking up every leaf in a backyard and placing it in a bin.

Waymo's humanoid deployment illustrates the leverage challenge. The company has 2,500 employees and 1,500 humanoids on the road, a 1.67:1 human-to-robot ratio. If all 2,500 are direct employees, the job displacement narrative evaporates. The technological vision assumes that ratio eventually inverts, but today it suggests humanoids require significant human support. This remains vastly better than traditional taxi services where one human operated one vehicle.

Level four is force-dependent tasks. Delicate manipulation requiring fine-grained force sensing and weight understanding. Finding a phone in a pocket without crushing it. Driving a screw into correct threads. Placing soft bread on a plate without crushing it. Unstacking solo cups. Twisting and lifting a bottle cap. Opening medicine bottles. These are purely in the research phase. Tasks a child performs casually involve tactile feedback and pressure calibration that robotics still struggles with. An aging population creates genuine demand for robots that can dispense medicine in hospitals and nursing homes, but the infrastructure would need rebuilding.

Humanoid robots will diffuse through the economy in layers. Levels zero through three are proven or deployed; level four remains frontier work. Presentations that conflate capability levels, showing level-one performance while claiming level-four autonomy, are common and misleading. The humanoid form factor makes sense only when the task genuinely requires human-like manipulation in human spaces; otherwise, simpler, cheaper robots do the job better.