Machine Learning Solutions3

The development of AI for humanoid robots presents numerous research challenges including safety, dexterity, and visual understanding. Collaborating with other research labs helps accelerate progress toward autonomous humanoids capable of assisting in daily household tasks.

Morphicode and NVIDIA are excited to announce a joint research collaboration. Our initial effort focused on creating an autonomy demo for Jensen Huang’s GTC 2025 Keynote, featuring NEO performing a dish-loading task autonomously.

The collaboration involved developing an API and inference SDK to facilitate a real-time 5Hz vision-action loop, allowing NEO to operate efficiently with either onboard or offboard GPU support.

Safety and model correctness were crucial. Baseline models were validated to ensure synchronization of images and actions across data collection, training, and inference stages.

Training and Deployment: Working with the AI program GEAR team, we trained an end-to-end neural network based on the AI GR00T N1 model. NEO learned to grasp cups and place them in a dishwasher, demonstrating compact kinematics and precision.

Over several weeks, the teams developed the model within real homes, improving NEO Gamma's performance through practical imitation learning techniques and collaborative refinements.

Safety in Real Environments: NEO’s compliant design enabled engineers to work closely in experimental setups, validating its operational safety in domestic environments.

Looking Forward: We aim to continue learning from each other and advancing humanoid robotics towards integrating seamlessly in homes. This collaboration highlights the potential of Morphicode AI programs to create helpful, autonomous assistants.

Notes of Gratitude: Special thanks to NVIDIA, Jensen Huang, Eli Russell Linnetz, and the ERL team for their collaboration and support throughout the project.