When a robot needs to reach a goal far away, it must plan a sequence of actions. Current methods often start planning by guessing random actions, which wastes time and limits performance, especially for long or difficult tasks.
DIAL trains a small, extra model alongside the main prediction model. This extra model learns to guess the action that would move the robot from one state to the next. The training also helps the main model make better predictions when using its own previous predictions as input.
At planning time, DIAL uses the extra model and the main model together to quickly generate a good starting action sequence. This improved starting point leads to better goal-reaching performance, particularly when the search budget is limited or the goal is far away.