World-Model Security Research Expands the Threat Surface Around Embodied AI
Researchers mapped attacks across sensing, world modelling, planning, control and human interaction in embodied-AI systems.
Researchers mapped attacks across sensing, world modelling, planning, control and human interaction in embodied-AI systems.
US reporting linked the restriction to surveillance, remote-control and critical-infrastructure concerns surrounding connected Chinese robotic systems.
Reporting clarified that previously authorized products remain available while newly introduced covered devices face the authorization barrier.
MERICS assessed China’s broad robot manufacturing base while noting continuing dependence on foreign AI compute and software ecosystems.
AP cited 2025 shipment evidence placing Unitree and AGIBOT among the largest global humanoid suppliers.
DLAM introduces temporal constraints intended to make latent robot actions more coherent across sequential control decisions.
The paper proposes converting failed actions into localized corrections for flow-matching vision-language-action policies.
Reporting highlighted EngineAI T800 humanoids rolling off a Zhengzhou production line as debate shifted toward industrial scale.
US startups warned that excluding affordable Chinese robot hardware could also slow domestic robotics experimentation and product development.
Chinese authorities criticized the US measure as protectionist and indicated they would defend domestic companies’ interests.