RESEARCH · RESEARCH · #517
AALT: composition-aware active imitation learning that maximizes start-goal connectivity
The paper introduces Adaptive Agents via Latent Topologies (AALT), an active imitation-learning method that selects demonstrations to maximize expected gains in start–goal connectivity by organizing demonstrations into a topology of latent hub states and identifying high-value bridge demonstrations. In a simulated UR5e ordered-retrieval domain with 72 tasks, AALT raised success from 42/72 to 72/72 using only 3 additional demonstrations (5 transitions); after 20 demonstrations, the best baseline averaged 88.6% success using 98 transitions.
KEY POINTS
- The paper introduces Adaptive Agents via Latent Topologies (AALT), an active imitation-learning method that selects demonstrations to maximize expected gains in start–goal connectivity by organizing demonstrations into a topology of latent hub states and identifying high-value bridge demonstrations.
- In a simulated UR5e ordered-retrieval domain with 72 tasks, AALT raised success from 42/72 to 72/72 using only 3 additional demonstrations (5 transitions); after 20 demonstrations, the best baseline averaged 88.6% success using 98 transitions.
- By explicitly valuing compositional, bridge demonstrations and tying this objective to information gain about reachability, AALT can dramatically reduce required expert demonstrations for multi-start-goal task families.
WHY IT MATTERS
By explicitly valuing compositional, bridge demonstrations and tying this objective to information gain about reachability, AALT can dramatically reduce required expert demonstrations for multi-start-goal task families.