simulated_robot

Set up the environment

Needed packages: Python, torch, gym, pybullet, tensorflow, stable_baselines, random, numpy, math, pickle

Data description(files in data folder)

Due to the space limit, we included the demonstrations here. Download data.zip. And unzip it here.

dis5.pkl: contains 5 expert trajs for 3DoF

normal48.pkl: contains 48 expert trajs for 7DoF

dis5ee.pkl: contains 5 expert trajs for 3DoF (ee_position only)

normal48ee.pkl: contains 48 expert trajs for 7DoF(ee_position only)

all_53.pkl: contains all 53 expert trajs(5+48)

dis5_3model: rl feasibility model for dis5.pkl

normal48_3model: rl feasibility model for normal48.pkl

weights_3.pkl: the generated weight(feasibility) for 53 trajs(5+48)

3_fea.pt: feasibility VAE model

3_normal.pt: no feasibility VAE model

How to create feasibility model

cd into pdenv directory, the file to create feasibility model is main_feasibility.py.

Important flags:

--env-name: Use 'feasibilitypanda-v0' as the environment name

--mode: Use 'dis' to create rl model for 3DoF robot arm, and use 'normal' to create rl model for 7DoF robot arm

Sample call:

python main_feasibility.py --env-name "feasibilitypanda-v0" --seed 3 --save_path "../data" --discount 1 --mode 'dis' --disdata '../data/dis5.pkl' --normaldata '../data/normal48.pkl'

The file to generate weights(feasibility) from the rl model is generate_weights.py, this will be used to create VAE model

Sample call:

python generate_weights.py

How to create VAE model

cd into SAIL directory

Sample call:

To train VAE model with feasibility:

python train_VAE.py --state-dim 3 --expert-traj-path ../data/all_53.pkl --weight True --weight-path ../data/weights_3.pkl --size-per-traj 4000 --output-path ../data/ --epoch 1 --seed 3

To train VAE model without feasibility(baseline):

python train_VAE.py --state-dim 3 --expert-traj-path ../data/all_53.pkl --weight False --size-per-traj 4000 --output-path ../data/ --epoch 1 --seed 3

Tips:

Disable or Enable GUI

gym_panda/gym_panda/envs/panda_env.py Line 19 and 20

p.connect(p.DIRECT) # disable GUI
p.connect(p.GUI) # Enable GUI

Window/Linux

This version is able to run directly in Windows, change path in panda.py, generate_weights.py, and panda_env.py to use in Linux