Chinese researchers have presented an agent framework that coordinates a humanoid robot with the surgical team over long sequences, incorporating three overlapping layers of safety.
Assisting a surgical procedure means following dozens of steps following a human process that is constantly evolving. With RoboAssist, a team of researchers proposes an agent-based framework that separates two representations: on the one hand, the estimated, and only partially observed, state of the human team's work; on the other hand, the sequence of tasks that can be performed by the robot.
As requests change, the system updates its estimate of workflow, context, and task dependencies, and then reschedules only the affected portion of the sequence. Compared to a full reschedule, this approach reduces plan update latency and model token consumption.
Three levels of security
Safety is based on a multi-stage architecture: preventive regulation of movements, reactive control when instruments pass through human contact, and independent supervision of the whole body during execution. All of this was demonstrated on a Unitree G1 humanoid in simulated scenarios of multi-stage surgical assistance, combining multimodal interaction, instrument handling, transport of medical equipment, navigation and safe handing over of objects.
Read the full article on the original source
Go to the source →