Dynamic Behavior Cloning With Temporal Feature Prediction: Enhancing Robotic Arm Manipulation in Moving Object Tasks

Abstract

In numerous real-world applications, the ability to accurately perceive and respond to dynamic changes in the environment, while also maintaining the flexibility to transfer learned skills across different tasks, is crucial for the effective operation of robotic arms. Behavior cloning is particularly promising in this context due to its data efficiency and strong task transferability, enabling robots to quickly adapt to new tasks by learning from demonstrations. However, traditional behavior cloning methods, which rely primarily on the observation and state information of the current frame to predict subsequent actions, fall short in dynamic contexts due to their static nature. To address this limitation, we propose Dynamic Behavior Cloning with Temporal Feature Prediction (DBC-TFP), which integrates with behavior cloning by leveraging historical frames to capture dynamic features crucial for predicting future scene images. This method uses a loss function based on the mean squared error (MSE) between the predicted future scene image and the ground truth counterpart, improving the model's accuracy in action prediction for dynamic scenarios. To evaluate our approach, we design a benchmark comprising eight task scenarios, including six foundational tasks and two advanced tasks. Experimental results on this benchmark demonstrate that DBC-TFP significantly improves the success rate of behavior cloning in dynamic scenarios compared to traditional behavior cloning methods.