The solution worked by assigning warehouse understanding and robot control to connected models instead of asking one fixed rule set to cover the entire task. ¹

WorkedDid not

Scene understanding

VLM on MEC interpreted images and task data

Fixed rules alone

Task planning

Generated subtasks, positions, and placement plans

Plans limited to known arrangements

Robot action

VLA used instructions and camera input

Recognition without action control

Training data

Real on-site labels for grasping and placement

Abstract data detached from operations

SoftBank trained the VLM with real on-site operational data labeled with grasping order and placement positions. ¹ The MEC-based model then generated object positions and placement plans even for unseen arrangements. ¹ This gave the upstream model a role in interpreting the work while leaving responsive physical control to the VLA on the Yaskawa robot. ¹

The architecture also kept the handoff explicit. The VLM understood the situation and generated subtasks from camera images and task-related data. ¹ The VLA used those instructions and camera input to generate safe and responsive commands. ¹ In combination, the models enabled flexible, highly accurate object picking and placement in changing warehouse conditions. ¹

SoftBank links VLM and VLA AIs for flexible warehouse picking and placement