Investigate discretizing Gripper Control (or data)

caio-freitas / GraphDiffusionImitate

Diffusion-based graph generative policies for imitation learning in robotics tasks 🧠🤖

MIT License

1 stars 1 forks source link

Investigate discretizing Gripper Control (or data) #79

Closed caio-freitas closed 6 months ago

caio-freitas commented 9 months ago

Using task-space setpoint positions for the gripper, it's possible that the grasping is poorly represented (would explain current performance). Try out simplifying gripper as open/close on the environment, and maybe on the dataset as well.

caio-freitas commented 9 months ago

Histogram of gripper values from the whole RobomimicGraphDataset looks like

Using -0.03 as threshold value for the gripper action seems reasonable