heewonpark12/v1_so101_stackcube

v1 — stackcube, no-core low-rank (SO-101)

Merged VLM delta: 0.17 GB per task.

Settings

cfg.load_moe            = False
cfg.num_images_in_input = 2                # top, wrist
cfg.task_suite_name     = "stackcube_so101"
cfg.unnorm_key          = "stackcube_so101"

SO-101 robot constants (prismatic/vla/constants.py) must be ACTION_DIM=6, PROPRIO_DIM=6, NUM_ACTIONS_CHUNK=20. The default LIBERO profile (7/8/8) fails to load the proprio projector.

action dim	6
proprio dim	6
action chunk	20
cameras	top, wrist (256x256)

Provenance

Compressed from ImKyungjin/mergevla-stackcube-so101; base VLM MergeVLA-Qwen-0.5B.