Domain-Adaptive Meta-Learning for One-Shot Imitation learning

with Pytorch implementation

1. This project refer to 2018:One-Shot Imitation from Observing Humans via Domain-Adaptive Meta-Learning;

2. The author's code repository:https://github.com/tianheyu927/mil;

3. The code implements two models with pytorch, simulatted pushing experiment and pushing experiment.

Model structure

DAML process illustration[DAML](https://arxiv.org/abs/1802.01557)

DAML network structure[DAML](https://arxiv.org/abs/1802.01557)

Environment

• Pytorch 1.3.1, conda with python 3.7.4
• numpy
• pillow

Usage

Simulating pushing

To download data, run

./scripts/get_data.sh

Set the model in main.py to model_sim, code in model_sim.py.

Simulated pushing is a simplified case that it uses robot demo and does not include state prediction Gym relevent code is not included, you have to write it yourself, please refer to https://github.com/tianheyu927/mil and check its code for detail.

Pushing

Set the model in main.py to model_pushing, code in file model_pushing.py, the default model parameters are set with the same values as in the paper.

collected data

the collected data is labeled with the number of objects shown in the images, the first number is the target.

a typical scene

How to use your own data?

Data should be prepared in accordance with what dataloader.py describes:

• image height/width/channel: the input images, corresponding to the input data;
• states: Robot states or configuration, which consists of the pose of the end-effector represented by the 3D position of 3 non-axis-aligned points on the gripper, please check the dataloader.py for correctly build your own dataset;
• action: groundtruth includes 3 linear velocities, 3 rotation velocities and 3 gripper pose for the loss computation;
• gripper_pose_min/max: the index in the action vector that indicates the pose of gripper, which is used for the gripper pose loss calculation;
• Manually adjust padding layer