Sber Robotics has launched the Green Challenge, a competition within the AI Journey 2026 event. Participants are tasked with developing or fine-tuning a control model for 'Green', the company's first humanoid robot, to perform mobile manipulation tasks in a digital simulator.
The model must process RGB images from three cameras, text instructions, and proprioceptive data to predict joint positions. These commands are executed by a Whole-Body Controller. The challenge focuses on Physical AI, requiring the model to handle states outside the training distribution and recover from execution errors.
Sber provides a comprehensive toolkit, including the NVIDIA Isaac Sim-based environment, the Green-VLA model with 1.3 billion parameters, and a dataset in LeRobot v2.1 format. The dataset contains 24,055 episodes, 72,165 video clips, and 15.17 million time steps, totaling approximately 140.5 hours of trajectories.
Data sources include scripted demonstrations, teleoperation, and synthetic Mimic expansions. Scripted data covers 15 tasks across five scenes. Participants may use the provided data or collect their own to improve model robustness against deviations that occur during autonomous task execution.
