Make grasp dataset in simulation

[graspnet]

Make grasp dataset in simulation

[Fang-Haoshu]

Hi, for using it in the simulation, you do not need to retrain the network. We directly used this repo and use the realsense model. The results is pretty good. Our model is robust in both realworld and simulation

Best

[elevenjiang1]

Hi, for using it in the simulation, you do not need to retrain the network. We directly used this repo and use the realsense model. The results is pretty good. Our model is robust in both realworld and simulation

Best

Thank you for reply,however the reason why I want to retrain it in simulation is:

1. Objects are difficult to get (though can buy from network)
2. To compare same network with different objects or even different network with different objects;This is aim to explore shape and grasp relationship in project like EGAD

By the way, we test graspnet in OCRTOC competition, it is not robust in some scenes.
For example in scene 2-2-3,only two grasp for a single bow

Best

[Fang-Haoshu]

I see. To generate data in simulation would be easy with 6D pose and any rendering tool. I would recommend pyrender.

The graspnet-baseline will produce dense prediction. For this case, I guess either the threshold is too high. or the coordination system is not set correctly.

In our updated version of the baseline, the performance is satisfactory and the demo is at: https://www.bilibili.com/video/BV17b4y1Z75f?zw It will also be available soon. Stay tuned!