Embodied-AI software lags robot hardware, says Unitree chief

Speaking at the World Robot Conference in Beijing on 20 August 2026, Unitree chief executive Wang Xingxing said software, not hardware, is the main bottleneck. He also described a goal for the future: a robot entering an unfamiliar home should complete about 80 per cent of tasks requested by natural-language instruction.

The remark is an important counterweight to humanoid demos: perception, planning, manipulation and recovery in unfamiliar environments remain unsolved.

The 80 per cent figure is an aspiration, not a proven benchmark. Task-level success rates, the safety case, intervention rates, maintenance costs and performance outside staged environments should form the evidence for any procurement.