My picking task is not picking with image or hybrid mode.

[RochMollero]

Hello,

I am currently trying the repo on a picking task in simulation. I have recorded 277 trajectories with teleoperation ) and with the low_dim version there are some good results ( the robot manages to pick objects that are positioned differently ).

However with the image or image version I have no luck : the robot goes 'around' the object but never 'takes it' (it happened once).

Where could this come from ? Could it be be cause the dataset is too small or might there be something else to check ?

For info I have two images, one attached to the gripper and the other one a global view of the bin. Both are fed RGB with 200x200 size and cropped at 180x180. I use the 9 current position targets as low_dim observation, and the predicted action is the n_action_steps values (including the current one).

Regards,
Roch