Sberbank has introduced GreenWAM, an AI model designed to control robots and predict environmental changes resulting from their actions. The system integrates world modeling with action planning, allowing it to forecast up to one minute of video based on robot movements.
The model was trained on data from 20 robot families, ranging from two-armed manipulators to anthropomorphic units. In the WorldArena 1.0 benchmark, GreenWAM ranked second among 128 models, outperforming solutions from NVIDIA, Stanford University, and AgiBot in geometry reproduction.
In the RoboTwin 2.0 test, GreenWAM achieved a 67.5% success rate, surpassing the 58.4% result of the π0.5 model. In the LIBERO-plus benchmark, which tests performance under varying conditions, the model completed 86.8% of tasks without additional training.
Testing on the proprietary humanoid robot Grine showed improved performance compared to NVIDIA’s Cosmos 3 Edge. Success rates in familiar conditions rose from 88.9% to 95.6%, while performance in new environments increased from 80% to 84%.
Developers also released an accelerated version of GreenWAM for real-time applications. This variant operates 12.3 times faster than the original, with the task success rate on the robot decreasing only slightly from 95.6% to 91.1%.
