This paper introduces Bi-ACT (Bilateral Control-Based Imitation Learning via Action Chunking with Transformer), a novel approach that integrates bilateral control principles with the Action Chunking with Transformer (ACT) model to enhance robotic control. The method aims to create a more robust and efficient control mechanism for autonomous manipulation tasks. Bi-ACT processes joint angles, angular velocities, forces, and images to predict the leader robot's actions, enabling the follower robot to perform nuanced and responsive maneuvers. The key contributions of the paper include: 1. **Novel Approach**: Bi-ACT combines bilateral control and ACT to handle both position and force information, enhancing the robot's adaptability and precision. 2. **Data Collection**: The system uses bilateral control to collect high-quality data, including position and force information, which is crucial for effective learning. 3. **Learning Architecture**: The model processes multimodal inputs (images, joint angles, angular velocities, and forces) and predicts the leader robot's actions over multiple time steps, achieving a frequency of 100Hz for fast and robust motion generation. 4. **Experimental Validation**: Extensive real-world experiments on pick-and-place and put-in-drawer tasks demonstrate the effectiveness of Bi-ACT, showing high accuracy and adaptability to various objects and environments. The paper concludes by outlining future directions for improving Bi-ACT, including enhancing robustness, integrating multimodal sensory inputs, and expanding its applicability across diverse robotic platforms.This paper introduces Bi-ACT (Bilateral Control-Based Imitation Learning via Action Chunking with Transformer), a novel approach that integrates bilateral control principles with the Action Chunking with Transformer (ACT) model to enhance robotic control. The method aims to create a more robust and efficient control mechanism for autonomous manipulation tasks. Bi-ACT processes joint angles, angular velocities, forces, and images to predict the leader robot's actions, enabling the follower robot to perform nuanced and responsive maneuvers. The key contributions of the paper include: 1. **Novel Approach**: Bi-ACT combines bilateral control and ACT to handle both position and force information, enhancing the robot's adaptability and precision. 2. **Data Collection**: The system uses bilateral control to collect high-quality data, including position and force information, which is crucial for effective learning. 3. **Learning Architecture**: The model processes multimodal inputs (images, joint angles, angular velocities, and forces) and predicts the leader robot's actions over multiple time steps, achieving a frequency of 100Hz for fast and robust motion generation. 4. **Experimental Validation**: Extensive real-world experiments on pick-and-place and put-in-drawer tasks demonstrate the effectiveness of Bi-ACT, showing high accuracy and adaptability to various objects and environments. The paper concludes by outlining future directions for improving Bi-ACT, including enhancing robustness, integrating multimodal sensory inputs, and expanding its applicability across diverse robotic platforms.