A team of researchers has upgraded robot contact detection from a simple two-way guess to a three-class model that can tell a human from a hard or soft object in real time.
The work used Franka Emika Panda robot arms and a custom dataset of contact interactions. Researchers trained three types of sequence models — LSTM, GRU, and Transformers — on time-series proprioceptive sensor data. The best model hit 91.11% accuracy during live testing. The team also found that a sliding window approach to preprocessing outperformed other strategies for this kind of time-series task.
In shared workspaces where robots and people operate side by side, identifying what a robot arm is actually touching matters for safety. Prior systems could only classify contacts as soft or hard — useful, but not enough when a human hand and a foam block feel similar. A three-class model that specifically flags human contact adds a layer of precision that safety systems need before any of this leaves the lab.
Ninety-one percent accuracy is a solid number on a controlled dataset; whether it survives messier real-world conditions — novel objects, varied contact angles, different body parts — is the question papers like this rarely get to answer.