In September 2026, frontier models such as OpenAI’s GPT-6 Astra and Anthropic’s Opus 5.5 are making the case for artificial general intelligence (AGI) increasingly hard to dismiss, at least for work done on computers. Whether that threshold has already been crossed remains contested. The question we are focused on is how soon comparable intelligence will arrive in the physical world.
There is no single agreed scientific definition of AGI. A widely cited starting point is Shane Legg and Marcus Hutter’s 2007 definition of intelligence:
“Intelligence measures an agent’s ability to achieve goals in a wide range of environments.”
Applying the familiar human-level interpretation of AGI, with reliability made explicit, we use the following adapted wording:
AGI: An artificial agent’s ability to achieve goals or reliably perform the full range of cognitive tasks that an average human can do, without failing in ways that would surprise a person given the same task.
Physical AI brings intelligence together with sensors, actuators and mechanical systems to perceive, reason about and act in the real world. At 120ft, we use the same wording with one change to describe physical general intelligence:
PGI: An artificial agent’s ability to achieve goals or reliably perform the full range of physical tasks that an average human can do, without failing in ways that would surprise a person given the same task.
This sets a higher bar than the robotics “ChatGPT moment” described by Unitree CEO Wang Xingxing following the company’s IPO in August 2026: a robot entering an unfamiliar home and completing around 80% of tasks through simple voice or text commands. PGI adds the breadth and reliability described above.
From our work with frontier labs and the latest robotics platforms, we believe PGI will arrive before the end of 2028. That is an ambitious forecast, including relative to Wang’s. The two hardest variables to judge are access to relevant training data and the precision with which robots can act. Recent progress on both helps explain our confidence.
Scaling laws for robotics
Several recent results suggest that robotics is developing some of the scaling behavior that drove progress in language models: increasing relevant training experience can produce predictable improvements.
Figure reports 56% complete-task success for Helix 2.5 across three behaviors in 30 unseen homes, compared with 9% without its Index pretraining. No additional training or adaptation was needed for those homes or objects. Figure also found that increasing pretraining data predictably reduced robot-action prediction error, and reports collecting roughly 35 minutes of human experience every second. The scaling result measures action prediction; the home trials show how that broader experience can translate into useful behavior.
Generalist’s GEN-1.5 points in the same direction. Across ten short tasks, the company reports 59% average success from a single demonstration supplied as input, without retraining. With five minutes of demonstrations per task, roughly 50 examples, and ten training updates, success rose to 83%. The reported standard deviations were 10 and 9 percentage points respectively. These results suggest that a sufficiently capable base model can make a new physical task much cheaper to learn.
General-purpose models are also becoming more capable robot controllers. In RoboCurve’s evaluation, GPT-6 Astra placed a block into a bowl in 19 of 20 trials, or 95%, compared with Claude Fable 5.1’s 8 of 20, or 40%, using the same robot hardware and agent framework. These are small samples on a simple task, but they show how much physical performance can change with the underlying model.
MIT’s Phillip Isola describes the emerging approach as robot-use agents: general-purpose models operating robots through tools, writing control code and learning new tasks with little or no robot-specific training. This is one reason we think the pace is easy to underestimate. Robotics can draw on advances in coding, vision and reasoning developed across the wider AI industry. Knowledge useful in the physical world does not have to come exclusively from robot demonstrations.
World models offer another route to more useful experience. World Labs’ real-to-sim-to-real research reports reconstructing physical tasks in simulation, training robot policies there and transferring them back to hardware. Spatial intelligence benefits when a robot can practise variations of a task without every attempt requiring someone to reset the physical scene.
AI is also starting to do parts of the work required to improve robots. Waddle Labs describes an agent collecting roughly a thousand LEGO pick-and-place trials overnight and then training a controller from that data. This is an early, practical route toward recursive intelligence: AI helping produce the experience and skills from which better capabilities can be built. Our 2028 forecast depends partly on this shortening of the learning cycle.
The final millimeters
The physical feedback loop is still a weakness in many visually driven systems. A robot may understand where a connector belongs and still struggle to insert it. Alignment, friction and resistance matter most at precisely the point where the camera can no longer tell the whole story. Touch, force sensing and fast corrections become essential.
The same RoboCurve evaluation makes this gap visible: Astra succeeded in only 2 of 20 puzzle-insertion trials. Getting an object to the right area and completing a precise physical interaction remain very different capabilities.
We expect progress to come from combining general reasoning with specialised skills and fast local control. A model can plan the sequence, write or help train the controller, and interpret failures; a faster feedback loop can handle contact and small corrections. Anthropic’s robotics research already shows how strongly performance depends on the interface between the model and the robot.
This is why we are optimistic about seeing a robot assemble an unfamiliar LEGO set from its instructions within two years. It would bring visual understanding, planning, precise manipulation and recovery together in a task anyone can understand. It would be a useful milestone on the way to PGI. For the full 2028 forecast to hold, those capabilities must transfer across the range of ordinary physical tasks.
The implications extend to recursive manufacturing: robots helping build and improve the equipment that produces more robots and other goods. Physical capability could then help expand the machinery through which AI acts on the world. That feedback loop can begin with human-level capability; it does not require superintelligence or a fully self-replicating factory.
We are not alone in expecting the next two years to matter. In Y Combinator’s discussion with Waddle Labs and RoboCurve, Jay Chooi describes an expectation among some frontier labs and robotics companies that general-purpose robots could arrive within two years or sooner. We share his concern that society is unprepared.
Our call is PGI before the end of 2028. That is a forecast about capability; widespread deployment will take longer. We explore the path from capability to adoption in The Humanoid Economy in 2027. But education, workforce planning and industrial investment already operate on longer timelines than the breakthrough we expect. Society is not preparing for general-purpose robotics in the next two years. We believe it should be.