Asymmetric On-Policy Distillation Learns In-Context Exploration
ASTEROID learns to explore with roughly 100 times fewer interactions than reinforcement-learning baselines, the authors report.
A student agent gathers its own experience with incomplete information, then receives action labels from an expert with privileged information. Training on those histories teaches the student to search while uncertain and act once hidden information is revealed. Across 11 environments, the authors report roughly twice the returns under fixed exploration budgets. Policies transferred from simulation to real-world robot manipulation without vision also increased success by a reported 40% over standard distillation.
ArXiv · AI/CL/LG's note
A student agent gathers its own experience with incomplete information, then receives action labels from an expert with privileged information. Training on those histories teaches the student to search while uncertain and act once hidden information is revealed. Across 11 environments, the authors report roughly twice the returns under fixed exploration budgets. Policies transferred from simulation to real-world robot manipulation without vision also increased success by a reported 40% over standard distillation.
ArXiv · AI/CL/LG's note
score 5