Dual-Stage Attention-Augmented ACT Model: Optimization and Validation for Bimanual Fine Manipulation Tasks

Traditional Action Chunking with Transformers (ACT) exhibit limitations in multi-step precision manipulation tasks (e.g., battery insertion, nylon cable tie threading). Their single-stage attention mechanism struggles to capture cross-temporal dependencies, leading to significant deterioration of action coherence during critical phases in later task stages. Concurrently, the static fusion approach for multimodal features restricts adaptability in dynamic contact scenarios. To address these challenges, this paper proposes a dual-stage attention mechanism: A global self-attention layer establishes long-term dependencies in action sequences, while a conditional cross-attention layer dynamically filters critical contextual features (including latent variables and visual cues). These cascaded operations enable closed-loop optimization from global planning to local calibration. The hierarchical architecture integrates three functional modules: enhanced initial action sequence modeling at the encoder side, adaptive multimodal fusion through cross-module interfaces, and optimized end-effector control at the decoder side. Experimental results demonstrate that the improved model achieves 8% higher success rate in Cube Transfer tasks and 14% improvement in bimanual insertion tasks, with notable robustness in high-precision scenarios. Ablation studies confirm the necessity of hierarchical components: Removing encoder attention induces cumulative planning errors, disabling cross-module attention causes multimodal feature misalignment, and eliminating decoder attention exacerbates control instability. This research provides an extensible sequence modeling paradigm for complex manipulation tasks.