According to CryptoSlate, a preprint report submitted by the PEAR team on November 13 shows that their model, trained after robotic task simulation, achieved 18 successes out of 30 grasp-and-block-alignment tests across different physical hardware entities. Both types of tasks had previously been practiced in simulation, and "simulation-only" is limited to the post-training phase of the robots in this study rather than the entire training of the base model. The results are reported by the authors and are still awaiting independent replication. Furthermore, there are limitations regarding prompt dependency and inference latency; therefore, it cannot be concluded that the model has mastered arbitrary real-world robotic tasks.