HOST: So if I'm building agents that handle several jobs in a row, why should I care about this?
EXPERT: You might need an agent to find the next object after the last job leaves it facing the wrong way. The authors studied that handoff in Minecraft.
HOST: Doesn't telling it the next job solve that?
EXPERT: Not if the object is out of sight. A goal can say what to find without showing where it is now.
HOST: So what does Ataka receive?
EXPERT: It's basically a picture where the object you're looking for is marked. The catch is that this training picture comes from another world, so the scenery doesn't line up perfectly with the route in the agent's world.
HOST: So what is it going to do with that picture?
EXPERT: It learns from player demonstrations to search, approach, and interact, and training also asks it to predict whether the object is visible and which of those states it's in.
HOST: So what did the tests actually show? Then what do you think is still missing?
EXPERT: They found that Ataka hit a 39 percent clean success rate on the mine task. That's in MineCraft, where it did the right thing without a wrong class interaction. But there was still a scripted system providing goals and handling some steps between tasks, and they haven't tested it with an integrated planner or a robot yet.