Researchers have built an imitation learning system that picks its own training examples based on how many tasks they unlock, not just how much they teach.
The method, called Adaptive Agents via Latent Topologies (AALT), comes from a new arXiv paper on active imitation learning. Instead of selecting demonstrations that maximize information gain about an expert's policy, as prior methods do, AALT looks for demonstrations that connect the most start-goal tasks at once. It organizes existing demos into a network of reusable behavior hubs and asks an expert to fill in the bridges most likely to unlock the largest number of unsolved tasks. In a simulated UR5e robot arm test with 72 ordered-retrieval tasks, AALT went from solving 42 tasks to all 72 using just three additional demonstrations, totaling five extra transitions.
That efficiency gap is the real story. The strongest baseline needed 20 demonstrations and 98 transitions to reach only 88.6 percent success, meaning AALT hit a better result with a fraction of the expert effort. For robotics teams, expert demonstrations are the expensive bottleneck, not compute, so a method that asks for fewer, smarter examples is worth more than one that just learns faster.
It's one simulated benchmark, not a warehouse floor. But the underlying idea, that some demonstrations are structurally worth more than others, is the kind of bookkeeping robotics training has needed for a while.