Active imitation learning reduces expert effort by letting the learner request the demonstrations it needs. Existing approaches usually select requests based on expected information gain about the expert policy, but in structured multi‑task domains the number of start‑goal task combinations grows exponentially while many solutions share reusable behaviors. Composable behaviors are therefore especially valuable because a single demonstration can enable many tasks at once. Prior methods do not explicitly account for this value when choosing demonstrations.
We introduce Adaptive Agents via Latent Topologies (AALT), which requests demonstrations that maximize the expected increase in start‑goal connectivity. This objective is formally equivalent to information gain about task reachability. AALT organizes existing demonstrations into a topology of latent hub states linked by learned behaviors, identifies high‑value bridge demonstrations that can unlock many tasks simultaneously, and grounds each bridge to an expert query. At inference time the system plans through the topology and conditions a diffusion policy on each successive hub transition.
In a simulated UR5e robot ordered‑retrieval domain with 72 tasks, AALT achieved 72/72 (100%) success using only three additional demonstrations (five transitions) beyond the initial dataset. The strongest baseline, after 20 demonstrations (98 transitions), averaged 88.6% success.
Review: AALT’s explicit modeling of composable demonstrations dramatically improves the efficiency of active imitation learning in large task sets, highlighting the power of a few high‑impact examples for robot manipulation.