AI Agents Learn Zero-Shot Imitation Without Human Data via Goal Inference
Source Posts on X
Cluster sources
Key ideas: 1⃣ agent imagines goals and practices reaching them (a variant of @silviupitis's MEGA) 2⃣ agent learns to infer intentions (i.e., goals) from behavior. Test time: agent sees some behavior, predicts which intention/goal caused that behavior, tries to reach that goal.
Natural limitation: #2 is inverse RL on the restricted hypothesis class of goal-reaching reward functions (⬇️variance,⬆️bias). You can construct edge cases where this approach fails. But on standard benchmark tasks, it outperforms a baseline that considers all reward functions.
Subtle detail: inverse RL for goals is not as simple as picking states the expert visited most/last. Eg, the 2nd place finisher of a race was probably trying to win. 👆a common problem in prior methods for zero shot imitation. 🛠️The fix: account for difficulty of each goal.