The man who sells some of the cheapest humanoid robots in the world thinks the hardest part of his business is about to be solved by someone else. Wang Xingxing, founder and CEO of Hangzhou-based Unitree Robotics, said Monday that OpenAI's recently released GPT-6 has convinced him a robotics breakthrough could arrive within one or two years. His most striking suggestion was a speculative one: that a GPT-7 model, released as early as 2027, might be able to control robots directly, without the layers of task-specific training that humanoid developers rely on today. It is a bold claim from an executive whose company has every commercial reason to want it to be true.
Wang's reasoning follows a pattern he has voiced before. Hardware, in his telling, has largely caught up. Unitree has spent a decade driving down the cost and complexity of actuators, joints, and legged locomotion, starting with its quadruped line and extending to humanoids such as the G1 and H1. What has lagged is the intelligence layer: the ability for a robot to understand an ambiguous instruction, perceive a cluttered environment, and execute a multi-step task without a human scripting each motion. Large language models have steadily eaten into that gap, and Wang sees GPT-6 as evidence the curve is steepening rather than flattening.
The distinction matters because of how the industry currently builds robot behavior. Most humanoid programs stitch together separate systems for vision, planning, and low-level motor control, then fine-tune on teleoperation data collected at considerable expense. Companies including Figure AI, Tesla, and Agility Robotics have each pursued variants of this stack, and several Chinese rivals, among them UBTech and AgiBot, are doing the same. A general-purpose foundation model that could collapse those layers would reduce the advantage of firms that have spent years accumulating proprietary demonstration data. Unitree, which has leaned on selling hardware widely and letting developers build on top, would be positioned to benefit from precisely that outcome.
Still, the distance between a capable language model and a reliable robot controller is large, and engineers will note the gap. Language models reason in tokens, while robots operate under hard real-time constraints, contact dynamics, and safety requirements that punish a single bad output with a dropped payload or a fallen machine. Latency alone is a serious obstacle when a balance loop must close in milliseconds. The more plausible near-term path is a hybrid in which a frontier model handles high-level reasoning and a smaller on-board policy handles execution. Wang's timeline of one to two years is aggressive against that backdrop, and the sharper observation is about incentives: a hardware maker is publicly betting its roadmap on a software partner it does not control, which may explain why Unitree has been so vocal about it.
What to Watch
Track whether OpenAI or any GPT-6 partner publishes embodied-control benchmarks or robotics-specific tooling over the next 90 days, since Wang's thesis depends on that evidence appearing. Watch Unitree's next G1 and H1 software releases and any announcements of foundation-model integrations or developer SDK updates, as well as its progress toward a public listing in China. Compare moves from Figure AI, Tesla's Optimus program, and AgiBot, whose own model announcements will show whether the industry is converging on Wang's view. Finally, monitor any 2027 release signals around a GPT-7 class model, because that date is the one Wang has effectively put on the calendar.
