Coding agents can now write and fix software by themselves, but robots face a harder problem: they must learn from physical actions, like grasping and moving objects, not just from error messages. A failed grasp might only make sense after seeing how it led to a later collision.
Researchers tested thousands of robot task attempts in simulation, varying the agent's model, the skills it had, the feedback it got, and the order of learning. They found that giving more detailed data about each moment of motion did not help much; what mattered was feedback that highlighted meaningful decisions.
The key finding was that agents often failed not from missing information, but from misinterpreting it, struggling with spatial reasoning, geometry, and translating errors into better motion plans. This suggests that future progress depends on designing smarter feedback and representations, not just bigger models.