Apple's HALO Hits 91% on Unitree G1; Incumbent Cuts 3,460 Jobs
Apple (苹果) researchers published HALO, a humanoid control system reporting a 91% average task success rate when simulated motion is transferred to previously unseen real rooms on Unitree's (宇树) G1 platform. The same day, a century-old industrial manufacturer disclosed 3,460 job cuts alongside a stated commitment to China's humanoid robot supply base.
The 91% figure matters less as a headline than as a protocol: evaluation ran on third-party hardware rather than a bespoke body, which makes the result comparable across platforms. The arXiv paper covers transfer to unfamiliar rooms, the failure mode that most often separates laboratory demonstrations from deployable control. If the number survives independent replication, control software begins to decouple from chassis design and the competitive axis shifts toward data and evaluation standards.
The 3,460 cuts accompany reallocation rather than retreat. Manufacturers holding legacy mechanical and drive-train capacity are repositioning toward embodied AI supply chains, where Chinese suppliers already hold cost and lead-time advantages. Capital released from mature lines tends to move toward components with defined specifications and repeat orders, placing actuators, reducers and sensing modules ahead of complete robot platforms in the near term.
速腾聚创（RoboSense）brings a Space Camera and digital LiDAR to CoRL 2026 with live 3D perception demonstrations, addressing the layer robot-learning pipelines depend on for state estimation. A separate result, in which a robot forms its own UV-sensitive sensor through a moth-inspired vascular channel, points at reducing integration burden by letting sensing emerge from structure instead of being fitted onto it.
Two observation points follow. First, whether the 91% success rate is reproduced on platforms other than the G1 under independent evaluation protocols. Second, whether capital freed by workforce reductions converts into component orders and standardized sensing specifications within the next two quarters.