What does HALO’s 91% success rate tell us?

  • Apple’s research team disclosed that its HALO system conducted 360 tests of individual tasks across 6 real-world environments that had not been scanned or mapped in advance, with an average success rate of 91%. This shows that robots can, to some extent, rely on autonomous control to complete individual tasks in real-world environments.
  • Results for combined tasks were significantly lower: in another real-world environment, 3 types of combined tasks were each tested 10 times, with an average success rate of 70%; the success rate for the task “bend down and pass an obstacle—walk to a target—push a box” was 50%. Execution stability still declines when moving from individual actions to continuous tasks.

What shortcomings limit the extension of success rates to more scenarios?

  • Data generation still involves human participation: new physical tasks require people to write prompts and design action keyframes, while some actions also draw on human videos or motion-capture footage. Therefore, the test results do not mean that the robot can already adapt by itself to any new task.
  • HALO currently relies mainly on vision and has difficulty directly sensing contact. When an object leaves the camera’s view or is occluded, performance on tasks involving substantial contact declines. Apple’s team has identified tactile perception as an area for future improvement, while data scarcity, labeling difficulties, and the complexity of physical environments also pose barriers to generalization.

Why are institutions focusing on “large-brain/small-brain integration”?

  • The “large brain” is responsible for cognition and decision-making, while the “small brain” handles motion control; only through their coordination can perception, decision-making, and execution form a closed loop. The report describes the industry’s evolution as moving from layered, collaborative architectures toward deeper integration, while also noting that the current integration of large AI models with humanoid robots remains shallow.
  • Dong Xiaobin of Industrial Securities believes that robot hardware processes have reached a relatively high level and that embodied-intelligence capabilities may become a development bottleneck; application scenarios remain relatively focused for now. If dedicated scenarios expand, training methods improve, and hardware continues to make breakthroughs, the accumulation of high-quality data may accelerate and drive applications toward general-purpose and household scenarios.

Are related companies’ developments at the same stage?

  • No. Joyson Electronics says it has launched components such as large-brain/small-brain controllers, but products related to North American customers are still progressing through sample submission, iterative upgrades, and mass-production work. JWIPC says customer orders have been secured in the field of robot large-brain/small-brain controllers. SigmaStar’s large-brain/small-brain SoCs for embodied intelligence, meanwhile, have begun entering tape-out in succession.
  • These developments relate to different stages for components, controllers, and chips. They reflect progress in products and projects across the industry chain, but sample submissions, orders, and tape-outs should not be treated as the same stage of commercialization.