Category: Electronics

Unitree founder and chairman Wang Xingxing told a business conference in Hangzhou on November 6 that intelligent security robots represent a promising commercial opening, citing strong overseas demand while conceding that humanoids are not yet in large-scale use.

Speaking at the eighth Zhejiang entrepreneurs’ conference, Wang presented demonstration video clips showing robots performing security and emergency tasks, such as restraining a simulated attacker, blocking an intruding vehicle and conducting emergency rescues. However, the address served as an outlook rather than a commercial rollout. Unitree announced no dedicated security models, commercial buyers or field deployments alongside the talk, leaving the robots’ actual performance during genuine emergencies unproven.

Addressing the state of the technology, Wang noted that humanoid machines worldwide have yet to enter large-scale trials or widespread adoption, remaining less efficient than conventional industrial automation. Unitree has pursued factory applications since 2020, testing units in its own production facility as well as with several automakers and automotive parts suppliers.

Inside predictable factory halls, robots trained on narrow, repetitive tasks can approach a 100 percent completion rate, Wang noted. Domestic living spaces present the opposite extreme, where unpredictable furniture layouts and varied human behaviour continue to block autonomous consumer deployment. Wang framed the ultimate technical goal as an embodied system capable of operating autonomously across unfamiliar physical environments without site-specific pre-programming.

Millimetre Disconnects in Embodied AI

Wang projected that within the next year or two, embodied systems might execute roughly 80 percent of tasks across unfamiliar settings relying solely on direct voice prompts. He identified that inflection threshold as the physical robotics counterpart to the emergence of consumer AI chatbots. Reaching that operational baseline remains blocked by hardware control limits.

The central technical hurdle lies in execution precision, Wang explained. Discrepancies between digital model instructions and mechanical motor responses routinely create positional errors of several millimetres. In delicate assembly, grasping or physical containment tasks, small deviations cause dropped tools, failed grips or mechanical collisions.

Compute Hardware Outweighs Algorithm Tweaks

Core advances across modern robotics now stem from automated cycles of data capture, scoring, model retraining and refinement rather than novel algorithm designs, Wang argued. Computational throughput and semiconductor capacity have therefore emerged as the decisive commercial constraints. Large-scale autonomous self-training workflows will demand massive hardware infrastructure before humanoid hardware can handle continuous real-world interaction.

Wang compared the current generation of robotics hardware to personal computers more than two decades ago. Early personal computers were widely dismissed as dedicated video gaming appliances because business software ecosystems had not matured, a pattern now mirrored in humanoid robots running staged performance routines instead of functional enterprise labour.

Factory Experience and Autonomous Targets

Commercial operators evaluating robotic site security or warehouse monitoring face clear operational trade-offs against conventional fixed surveillance. Automated cameras and basic teleoperated patrol rovers carry fractional hardware costs, proven reliability profiles and minimal liability risks compared with dynamic bipeds or quadrupeds attempting physical interventions. For facility managers, deploying unproven autonomous machines to intercept moving vehicles or restrain intruders risks substantial equipment damage, premises liability and erratic mechanical failures.

Unitree’s trajectory since its 2020 entry into automotive manufacturing facilities demonstrates that repeatable, enclosed manufacturing delivers actual runtime data, whereas multi-terrain security patrols demand open-world adaptation that current spatial models cannot guarantee. While factory pilot programmes with automotive manufacturers supplied training cycles, expanding into international commercial facilities requires crossing the millimetre gap between digital commands and physical joint responses.

The benchmark to track remains Wang’s 80 percent autonomous voice-task execution threshold across unmapped facilities over the next two years. Until robotic actuators eliminate minor millimetre positioning errors under real-world conditions, commercial deployments will stay restricted to controlled manufacturing cells rather than open security patrols.