aiqbench analysis
Unitree's founder is making a self-reported claim about embodied AI capability without published benchmarks, task definitions, or a method for measuring the "80%" figures. The more concrete point is his view that the bottleneck is mapping large-model outputs onto physical robot hardware, not high-level planning or navigation. Readers should treat the year-end timeline for household tasks as an aspiration until independent evaluations appear.
Unitree's Wang Xingxing: Embodied AI Has Passed the GPT Moment; Biggest Challenge is Aligning Models with Real Robots
36Kr reports that at the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference, Unitree founder Wang Xingxing reiterated his assessment of embodied AI's 'GPT moment,' stating that robots can now complete approximately 80% of major tasks in about 80% of unfamiliar scenarios via voice or text commands. He noted that while large models already demonstrate strong capabilities in robot planning and navigation, the primary challenge remains aligning AI model inputs and outputs with real-world robots. Wang expects that, if progress continues quickly, robots could soon follow instructions to perform tasks like tidying a room by year-end or next year.
Read the full article at 36KrRelated coverage
Bringing Physical AI to Real Industrial Conditions: Chenjing Technology Empowers Steel Mill Intelligent Inspection
36Kr · Oct 22, 2026
Wang Xingxing of Unitree Robotics: Physical AI Shows Signs of Being Disrupted
36Kr · Oct 22, 2026
Wang Xingxing of Unitree Robotics: My Biggest Mistake Was Not Believing Enough in AI
36Kr · Oct 12, 2026