One-Shot Visual Imitation Learning via Meta-Learning

A TensorFlow implementation of the two papers One-Shot Visual Imitation Learning via Meta-Learning (Finn*, Yu* et al., 2017) and One-Shot Imitation from Observing Humans via Domain-Adaptive Meta-Learning (Yu*, Finn* et al., 2018). Here are the instructions to run our experiments shown in the paper.

First clone the fork of the gym repo found here, and following the instructions there to install gym. Switch to branch mil.

Then go to the mil directory and run ./scripts/get_data.sh to download the data.

After downloading the data, training and testing scripts for MIL are available in scripts/.

UPDATE (7/3/2018): to run the experiment with learned temporal loss as in the One-Shot Imitation from Observing Humans via Domain-Adaptive Meta-Learning paper, take a look at scripts/run_sim_push_video_only.sh script.

Note: The code only includes the simulated experiments.