Humanoid Robots at Human-Level Capability
A general-purpose humanoid robot can perform the majority of physical tasks a human worker can do in an unstructured real-world environment — not a narrow factory task or scripted demo, but flexible dexterity, locomotion, and object manipulation across settings it was not specifically trained for.
- Median year
- 2033
- P10 – P90 range
- 2028 – 2045
- Probability ever occurs
- 95%
- Last reviewed
- June 2026
Humanoid robots can do most of what a human body can do at work. Labour markets, logistics, elder care, construction, and domestic work face structural disruption. The bottleneck on physical production shifts from human time to capital and energy.
Humanoid robots remain useful but narrow — reliable within specific trained tasks but brittle outside them. The gap between controlled demo and unstructured real-world deployment proves harder to close than expected, and human physical labour remains essential through this window.
Where things stand
Humanoid robotics has moved faster in the 2020s than most observers expected — but “faster than expected” and “solved” are not the same thing. The gap between a polished demo and reliable operation across the messy, unpredictable environments humans work in every day remains large.
The current generation of platforms:
- Figure AI (Figure 02): partnered with BMW for factory work; demonstrated conversational task instruction using OpenAI models; raised $675M in 2024 at a $2.6B valuation
- Tesla Optimus: Elon Musk claims thousands of Optimus units working in Tesla factories by end of 2025, with external sales targeted for 2026; independent verification is limited
- Boston Dynamics Atlas: pivoted from hydraulic to electric in 2024; optimised for warehouse and industrial tasks; the most mature locomotion of any platform
- 1X Technologies (backed by OpenAI): focusing on domestic and care environments rather than factories
- Agility Robotics (Digit): operating in Amazon fulfilment centres doing tote-moving tasks — one of the first commercially deployed humanoid systems
The critical unsolved problems are not mechanical but cognitive-physical: generalising manipulation to objects and configurations the robot has never seen, recovering gracefully from failure, and operating safely alongside humans without extensive environmental modification. Current systems handle constrained tasks reliably but degrade sharply outside their training distribution.
The distinction from AGI is important. A software system could achieve AGI-level reasoning without any robotic capability. Conversely, a humanoid robot capable of general physical work might rely on relatively narrow AI for perception and motor control — the two milestones can arrive independently and likely will. What makes humanoid robotics a distinct event is that it requires solving physical embodiment: real-time sensor fusion, whole-body control, and manipulation of the infinite variety of objects in the physical world.
The reference year of 2033 reflects a central estimate consistent with the current pace of hardware and AI progress. The 2028 optimistic bound reflects Tesla’s stated ambitions and the possibility that current foundation-model approaches generalise faster than expected. The 2045 bound reflects the historical pattern of robotics timelines running long — the “unstructured environment” problem has resisted solution for decades, and current demos take place in carefully prepared settings.