Downloads · 30 days
0
cosmos3-aloha-clothfold-fd-lora
ZoreAnuj/cosmos3-aloha-clothfold-fd-lora
cosmos3-aloha-clothfold-fd-lora is a machine learning model from ZoreAnuj. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
LoRA finetune of NVIDIA Cosmos3-Nano forward-dynamics on qcez/foldingcloth314infer-200 — policy inference rollouts (9 success + 137 fail episodes) on a bimanual ALOHA. Trained on rollouts including failures so the wor…
Access
Public
Updated Jun 6, 2026
Repo size
180 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors123 MB · 82%
At a glance
- Access
- Public
- Created
- Jun 6, 2026
- Updated
- Jun 6, 2026
- SHA
- 9a30a2e4
Model card
From the Hugging Face model README
Cosmos 3 Nano — ALOHA Cloth-Folding FD LoRA (eval-focused, success+fail rollouts)
LoRA finetune of NVIDIA Cosmos3-Nano forward-dynamics on
qcez/folding_cloth_3_14_infer-200 —
policy inference rollouts (9 success + 137 fail episodes) on a bimanual ALOHA.
Trained on rollouts including failures so the world model can render failure modes —
built for policy evaluation (test-before-rig), not just demo replay.
- Base:
nvidia/Cosmos3-Nano; generator LoRA rank 16/alpha 32 onq,k,v,o_proj_moe_gen(~15.3M params). - Domain:
robomind-franka-dual(id 12), 20-D dual-arm[pos3·rot6d6·grip1]×2. - Actions: 14-D ALOHA joints → MuJoCo FK (menagerie
aloha.xml,left/right grippersites) → backward_framewise rot6d deltas; q01/q99 normalized (161,573 transitions). - Vision: images-in-parquet (3 cams: cam_high + 2 wrists) → concat_view, res 720, fps 50.
- Holdout: last episode file of each subset (1 success + 1 fail) excluded from training for eval.
iter_XXXXXXXXX/adapter.safetensors = LoRA adapters. code/ = loader (FK + parquet images),
datapacker, experiment config, normalizer, training TOML.
- Created
- Jun 6, 2026
- Updated
- Jun 6, 2026