but the performance don't well

you can find fetch and environment class in here

DDPG algorithm in here

main function entrance in here

we have trained a network which can get xyz in simulate(gazebo) for img with deep.

there have a fetch_gazebo world (including obstacle model, fetch and the world)

if you are ubuntu you can try

sudo apt-get install ros-indigo-fetch_*

then you will install fetch package in /opt/ros/indigo/share/...(i copy the package from here);

python demo.py

Except put your data and model in GPU, you need to using follow coda:

device_ids = [0, 1]
net = Net().cuda(device_ids[0])
net = nn.DataParallel(net, device_ids=device_ids)
optimizer = torch.optim.Adam(eval_net.parameters(), lr=LR) 
optimizer = nn.DataParallel(optimizer, device_ids=device_ids)
optimizer.module.step()