search

hyperplane-lab / RLAfford

RLAfford: End-to-End Affordance Learning for Robotic Manipulation, ICRA 2023
https://sites.google.com/view/rlafford/
88 stars 8 forks source link

Bug with door index? #8

Closed Xingyu-Lin closed 5 months ago

Xingyu-Lin commented 1 year ago

https://github.com/hyperplane-lab/RLAfford/blob/a0e650eeb79e2b62395e787b9ff131118109903e/MARL_Module/envs/tasks/franka_cabinet.py#LL144C21-L144C21

This line finds the cabinet index based on the first environment. However, self.cabinet_rig_name varies across environments. Is this a bug? I have met situations where index returns -1

Xingyu-Lin commented 1 year ago

Tentatively solves the issue by changing this line to 

                if id == 0:
                    self.cabinet_rig_name = list(rig_dict.keys())[1]

But maybe the author can provide more explanation and suggest better solution

Xingyu-Lin commented 1 year ago

I also wonder how much this changes the whole benchmark? Seems like this function is closely related to the reward function, which affects both training and evaluation metric

boshi-an commented 1 year ago

We have modified the dataset so that all modles only have 2 rigid bodies, one is fixed with the base, one is the moving part. We believe 1 will always be index of the moving part. Concerning how you had situations where index returns -1, we would like to further investigate this issue. Could you provide the case how the index returns -1?