Sample-Efficient On-Policy Imitation Learning from Observations