How WIRobotics is teaching humanoid robots to use human tools with AWS and NVIDIA

This article describes how WIRobotics used AWS and NVIDIA to teach the ALLEX humanoid robot to perform contact-rich manipulation tasks like using a power drill.

• ALLEX combines dexterous hands with 48-degree-of-freedom upper body and whole-body force sensing for responsive physical interaction
• WIRobotics built a cloud-native training workflow on AWS using EC2 g7e.48xlarge instances with 16 NVIDIA RTX PRO 6000 Blackwell GPUs
• Fine-tuned NVIDIA's GR00T Vision-Language-Action foundation model on synchronized teleoperation data from three camera views and 44-dimensional robot state
• Achieved 100% success rate on drill grasping and 50% on drill-to-bolt alignment after six-week engagement
• Data quality proved more valuable than volume; longer action horizons and additional denoising steps improved performance
• Developed repeatable workflow and VLA fine-tuning guidelines for teaching future manipulation skills

The project demonstrates how foundation models enable robots to learn complex tool-use tasks in human environments without task-specific engineering.