dg5f_bluecable60_sparsh_ac90

DexWild (dit_policy) action-chunking Transformer for a DG-5F dexterous hand + UR5e arm blue-cable manipulation task. The two GelSight tactile cameras are encoded by a frozen SPARSH (facebook/sparsh-dino-base) tactile encoder; the two RealSense RGB cameras use a vit_base backbone.

Trained on Kaz55/dg5f_ur5e_bluecable60ep_h5 (60 episodes). This repo contains the final checkpoint at 50,000 iterations.

Files

File	Description
dg5f_bluecable60_sparsh_ac90_50000.pth	Model weights @ 50k iters (≈3.5 GB)
agent_config.yaml	Agent architecture config (needed to rebuild the model)
exp_config.yaml	Full experiment config
hydra/config.yaml, hydra/overrides.yaml	Resolved Hydra config + CLI overrides
combined_ac_norm.json, combined_state_norm.json	Action / state normalization stats (required for inference)
train_data_bluecable60_ac_norm.json, train_data_bluecable60_state_norm.json	Per-dataset norm stats

The .pth alone is not enough to run the policy — you also need agent_config.yaml and the *_norm.json normalization statistics.

Model / task specs

• Observations: state 26-d (UR5e 6 + DG-5F 20), 4 cameras (realsense, realsense2 → RGB vit_base; gelsight1, gelsight2 → SPARSH, cam indices [2, 3]).
• Actions: 26-d (UR5e gello 6 + DG-5F cmd 20), joint-absolute.
• Action chunk (ac_chunk): 90
• Transformer: d_model 768, nhead 8, enc 4 / dec 6 layers, ffn 3200, gelu, dropout 0.1.
• SPARSH: sparsh-dino-base, in_chans=6 (2 bg-subtracted frames), pooling=register, frozen.

Training

Framework	DexWild dit_policy (data4robotics), accelerate bf16, single GPU
Agent / task	agent=transformer_sparsh, task=rdm_dg5f_sparsh
Iterations	50,000 (save_freq/keep_freq=10,000)
Batch size	32
LR / sched	1e-4, trainer=bc_cos_sched (cosine)
RGB backbone	vit_base (SOUP_1M_DH.pth)
use_lang / use_tactile (uSkin)	false / false (GelSight routed through SPARSH, not uSkin)

Intended use

Research / imitation-learning baseline for the DG-5F + UR5e blue-cable task. Not a general-purpose robot policy.