dg5f_speedtorque_sparsh_ac90
DexWild action-chunking Transformer trained on the DG-5F + UR5e
bluecablespeedtorque dataset.
- Task:
rdm_dg5f_sparsh (obs=26, act=26, joint_abs)
- Action chunk: 90
- Training steps: 100000
- Backbone: RealSense = RGB vit_base, GelSight = frozen SPARSH (sparsh-dino-base)
Files: final checkpoint dg5f_speedtorque_sparsh_ac90_100000.pth, agent_config.yaml, exp_config.yaml,
norm-stats JSONs, and hydra/ config.
Inference dependency (SPARSH)
The two GelSight cameras are encoded by a frozen facebook/sparsh-dino-base tactile encoder. The saved .pth does not contain the SPARSH weights, so to run this checkpoint you also need: (1) the tactile_ssl package + this fork's data4robotics/models/sparsh.py; (2) the SPARSH weights facebook/sparsh-dino-base (dino_vitbase.ckpt); (3) XFORMERS_DISABLED=1. See the RealSense RGB backbone (SOUP_1M_DH.pth) note in the agent config.
