X-G1

In this 2-day hackathon, we built a rapid-iteration pipeline for the Unitree G1 humanoid, moving from manual demonstration to autonomous policy testing. Technical Implementation: Teleoperation: We integrated Meta Quest 3 with MuJoCo, using NVIDIA Sonic to achieve low-latency control. This allowed us to manually complete complex tasks, including walking to tables and performing pick-and-place maneuvers with beverages and apples. Data Strategy with DeepLake: To handle high-throughput training, we used DeepLake to store and stream Lightwheel’s G1 beverage organization data. The efficient tensor storage provided the fast I/O necessary to fine-tune models within our tight time constraints. Policy Fine-Tuning: We fine-tuned NVIDIA GR00T on our collected data. Since Sonic’s fine-tuning features are not yet released, we used it primarily for high-fidelity data collection while running autonomous inference through GR00T. Diagnostics with Nomadic: To understand why our fine-tuned agent struggled with specific task instructions, we used Nomadic AI. This diagnostic layer allowed us to pinpoint failure modes in the model's reasoning, creating a clear path for future performance improvements. By combining immersive teleop with robust MLOps tools, we've demonstrated a fast-track workflow for humanoid robot learning. Note :: Forgot to add team members Sagar Patel :::sagarp3199@gmail.com and Malika Rakhimova :: malikakxan16@gmail.com please add as team member