Collaborative robots: a model learns to reject unfamiliar movements

A study combines spatial and temporal analysis to identify human actions, whilst flagging those that fall outside the learned categories.

Published on September 11, a study proposes a system of recognition of human actions for workspaces shared with robots. The model combines a spatial convolutive network, a temporal analysis and a mechanism that rejects actions considered unknown.

The authors evaluate NTU RGB+D 120 by reserving certain classes of actions to test this rejection. They include 66 frames per second on Jetson Orin NX. The stake is to avoid that a new gesture is automatically interpreted as a known action. These results on a data set are not valid for industrial safety certification.