You can deploy this model and start earning money today!

Kaz55/
dexwild-dit-dg5f-ur5e-newcable-v5v6

Name: dexwild-dit-dg5f-ur5e-newcable-v5v6
Availability: InStock

Views

No views yet

Kaz55/dexwild-dit-dg5f-ur5e-newcable-v5v6

DiT (diffusion transformer) policy trained with the DexWild dit_policy codebase on Kaz55/dg5f_ur5e_newcable_v5v6_h5.

Setup


robot	UR5e (6 dof) + DG-5F right hand (20 dof)
state / action	26 dims each — `ur_arm_joint_states`+`dg5f_right_joint_states` / `ur_gello_joint_position`+`dg5f_right_joint_cmd`
action space	`joint_abs`
cameras	realsense, realsense2, gelsight1, gelsight2
vision encoder	ViT-base per camera (`share_cam_features=False`), init from `SOUP_1M_DH.pth`
action chunk	60
iterations	200000
batch size	32
data	180 episodes / 208,933 transitions @ 30 Hz

The two GelSight streams are their own cameras with their own ViT weights, so tactile features are never forced through the RGB scene encoders.

Files

dg5f_newcable_v5v6_dit_ac60_200000.pth
agent_config.yaml
exp_config.yaml
combined_ac_norm.json
train_data_newcable_v5v6_ac_norm.json
train_data_newcable_v5v6_state_norm.json
combined_state_norm.json

agent_config.yaml / exp_config.yaml are the resolved hydra configs the run used; serve_policy.py in the training repo consumes the checkpoint directly.