World Models Receive a Security Threat Model
Researchers map cyber-physical attacks across sensing, world modelling, planning, control and human interaction.
Researchers map cyber-physical attacks across sensing, world modelling, planning, control and human interaction.
Action-conditioned predictive models are used to improve generalizable embodied decision-making.
RynnWorld-Teleop combines action conditioning, depth awareness and skeletal structure for egocentric teleoperation prediction.
A controlled study found prediction objectives determine whether robot world models retain mass, stiffness, drag, and force information.
Researchers map attacks across sensing, world modelling, planning, execution, feedback and long-term adaptation.
Researchers mapped attacks across sensing, world modelling, planning, control and human interaction in embodied-AI systems.
An exploratory mobile-robot study reports spatial memory and action consequence reasoning from a general multimodal model.