Motus2 reaches 84% on real-robot tasks

ShengShu Technology’s Motus2 world model reports an 84% average success rate across five real-robot dexterous-manipulation tasks, using 20 trials per task, in the team’s paper described by the launch release. Adding model-based reinforcement learning and inference-time planning lifted a two-task average from 65% to 75%, while more than 100 hours of robot-domain trajectories moved the five-task result from 51% to 84%. Robotics teams should treat this as a vendor-reported lab benchmark and test repeatability on their own hardware before using the archive’s physical-AI industrialization clock to set a deployment date.