Visuotactile Diffusion Policy: Automated Failure Recovery in Assistive Tasks with Tactile Manipulation Using Imitation Learning

Imitation learning is a powerful technique for teaching autonomous agents a variety of tasks. However, many imitation learning algorithms suffer from the fundamental problem of error propagation. Typically, imitation learning is framed as a supervised learning problem; however, during live operation, the agent operates in a state space generated by its own actions instead of expert demonstrations. This co-variate shift can lead to agents either failing to complete designated tasks, taking dangerous actions which can lead to damage or harm, or simply copying expert behavior without completing the task (copycat problem). The problems of co-variate shift and expert copying is especially important in safety-critical environments, such as assistive robotics, where simple errors can bear high costs. While algorithmic solutions exist for this problem, these often rely on constraining the agent policy or simply improving decision making without resolving the issue of error propagation. To address these challenges, we present an efficient solution for resolving error propagation by introducing the tactile modality in fine-grained grasping and manipulation tasks. To this end, we present Visuotactile Diffusion Policy, a policy learning framework which allows for automated failure recovery. The purpose of this study was to explore an out-of-the-box technique for preventing co-variate shift and the copycat problem, especially in grasping tasks for assistive robotic systems. The primary contribution of this study was to develop a robotic system and policy learning framework capable of automated failure recovery in grasping tasks. Along with this, we demonstrate how tactile sensing can lead to more robust robotic control policies and provide a generalizable solution for co-variate shift in robotic control tasks.