Sim-to-Real Acceleration Prediction Using Deep Learning for Shock Avoidance in Mobile Robots

抄録

Mobile robots operating on uneven terrain face shocks and vibrations that jeopardize stability and sensor reliability. To enhance safety, this study proposes a deep learning framework to predict the maximum resultant acceleration within a short future window. The model processes multimodal time-series data, including Inertial Measurement Unit (IMU) data, velocity, command velocity, and depth images. A Convolutional Neural Network (CNN) extracts spatial terrain features, while a Long Short-Term Memory (LSTM) network captures temporal dynamics to infer future physical loads.
Experiments were conducted using datasets generated in Isaac Sim across slopes and steps. The proposed model demonstrated real-time capability with an average processing time of 1.13 ms. Quantitatively, the model yielded a Root Mean Squared Error (RMSE) of 3.71 m/s² for a prediction horizon of 0.5 seconds. In comparison, a baseline model that directly propagates the most recent maximum acceleration to the future yielded an RMSE of 6.93 m/s², indicating that the proposed method significantly improves prediction performance. These results confirm the method's effectiveness in anticipating impacts, supporting its application in risk-adaptive motion planning for autonomous navigation.