Transformer-Based Sequence Modeling with Action Discretization for Robotic Grasping

The powerful representational and sequential modeling capabilities of Transformer have been validated in various fields such as natural language processing. However, how to better address sequential decision-making problems like robotic grasping through the Transformer architecture and achieve autonomous control of robotic grasping remains a longstanding challenge. Models that directly output continuous decision actions for robotic arms often rely on mean squared error (MSE) loss for training, which is unable to effectively handle the randomness and diversity of actions. This limitation leads to the generation of relatively singular actions in certain grasping scenarios, reducing the success rate of grasping. To address this issue, this paper proposes a Transformer-based method for discretizing robotic arm actions, where actions are discretized by dimension to enable autoregressive modeling. Experimental results demonstrate that this method effectively improves the grasping performance of robotic arms in simulation environments.