Slip Detection for Robotic Grasping with LSTM and TCN: Edge AI Evaluation on Embedded Platforms

Abstract Reliable object grasping is essential for robotic manipulation in real-world environments, such as manufacturing and assembly. Undetected slips can lead to task failure, while effective slip detection enables robots to adapt their grip in real time, improving robustness and safety. In earlier work, we showed that Long Short-Term Memory (LSTM) networks can be deployed on edge devices for real-time slip detection, but that study used a manually labelled dataset and reported only accuracy and F 1 scores, which limited comparisons of model–hardware combinations. In this work, we collect a large slip detection dataset using a Fin Ray-inspired soft gripper with tactile Force Sensing Linear Potentiometers (FSLPs), a Time-of-Flight (ToF) sensor, and an encoder, to generate slip and no-slip labels from automatically estimated ArUco marker poses. We design and evaluate lightweight LSTM and Temporal Convolutional Network (TCN) architectures trained on sensor sequences and compare their performance on a held-out test set. We then deploy the selected models on several embedded platforms, including Nvidia Jetson AGX Orin, Google Coral Dev Board Micro, ESP32, and Arduino Nicla Sense ME, and benchmark accuracy, F 1 score, average inference time, and estimated average energy per inference. The results show that accurate slip detection is feasible even on constrained microcontroller- class hardware, and provide a more comprehensive comparison of model and hardware choices than previous work.