Модель Pan-1 от Pantograph продемонстрировала способность переносить навыки из Minecraft на реальных роботов

Many users praised Pantograph's Pan-1 RL model for its versatile Minecraft task performance and robotics potential, while others dismissed its novelty compared to open-source mods and doubted the claims due to prior RL hype cycles. Based on 42 visible X reactions from 108 accounts. The model transfers virtual training behaviors toward physical robotics systems. this one makes more sense BC-wise: it does the same nervous / bored tic that minecraft youtubers do while waiting for this grass block to grow. partially breaks / pokes at the slow block One of the dreams of RL is to train models that can quickly learn how to act in any RL environment, even ones they’ve never seen before. This new model is a step in that direction: https://twitter.com/pantographPBC/status/2077423710617976897 one obvious thing to try here is can you map other stuff into the goal image space? like photos. or text? can we glue an llm to this thing? either post hoc or pretrained jointly... train this + a few T of webtext intermixed? (lets plays are narrated...) Pan-1 can accomplish a pretty wide range of visual goals in Minecraft. Now, Minecraft isn’t every RL environment, but it’s pretty open-ended. What I think is more interesting than the performance we got is how general the training method was. this is insane. IT WENT BACK FOR THE HORSE. how did it learn to do that offline?? https://twitter.com/pantographPBC/status/2077423710617976897 I sense there is a lot more interesting research to come from the quiet team at Pantograph https://twitter.com/pantographpbc/status/2077423710617976897 insane to think 1) this is a 4B model and 2) this method should translate almost directly to learning from real world video... gpt-2 moment Many users praised Pantograph's Pan-1 RL model for its versatile Minecraft task performance and robotics potential, while others dismissed its novelty compared to open-source mods and doubted the claims due to prior RL hype cycles. Based on 42 visible X reactions from 108 accounts.