Amazon Robotics’ ViTa-Zero solves key robotics challenge

Amazon Robotics has developed ViTa-Zero, a novel zero-shot visuotactile pose estimation framework designed to address the challenge of accurately determining object poses during robotic manipulation tasks. Traditional visuotactile approaches have struggled with generalization due to limited visuotactile data. ViTa-Zero innovates by using a visual model backbone combined with feasibility checks and test-time optimization grounded in physical constraints derived from tactile and proprioceptive sensor data. The system models the interaction between the robot gripper and object as a spring–mass system, where tactile sensors provide attractive forces and proprioception generates repulsive forces, enabling more robust pose estimation. The framework was validated on real-world robotic setups across various manipulation scenarios such as grasping, object picking, and bimanual handover. Experimental results demonstrated significant improvements over purely visual models, including a 55% increase in area under the curve (AUC) for ADD-S, a 60% increase in ADD, and an 80% reduction in