sadjava
/

smolvla-video-expert-libero90-s1000

Model card Files Files and versions

SmolVLA video expert · libero_90 · seed 1000

Action-free residual video pretrain of the SmolVLA action expert on RGB from nvidia/libero_90 (no action / state labels). Predicts SigLIP residual x_{t+Δ} − x_t (Δ=8). Init: lerobot/smolvla_base. Full action-expert weights (no LoRA).

Use as --policy.path for few-shot LIBERO-Goal FT (OUT_SUFFIX=_frombase_vpre or _frombase_vlora).

Files:

pretrained_model/ — config + model.safetensors (merged weights)
video_expert.json — training hyperparameters (if uploaded)

Downloads last month: 14

Safetensors

Model size

0.5B params

Tensor type

F32

·

BF16

·

Video Preview

loading

Collection including sadjava/smolvla-video-expert-libero90-s1000

libero_seen_pretrain

Seen-suite (libero_90) pretraining for LIBERO-Goal few-shot: SmolVLA action BC + action-free latent dynamics (IDM/FDM). • 5 items • Updated 7 days ago