Robot policy checkpoints
| folder | model | file |
|---|---|---|
vla_1f/ |
SmolVLA fine-tune, 1 observation frame. final = 81 epochs (20k steps); epNNNN = intermediate |
model.safetensors, config.json (lerobot 0.3.3 SmolVLA format) |
vla_3f/ |
SmolVLA fine-tune, 3 observation frames with a temporal-fusion conv (custom policy class smolvla_mf) |
same |
residual_s1/ |
event-based residual policy on top of vla_3f/final |
model.pt (state dict), args.json (normalization statistics), model_step*.pt |
vla_3f_t123/ |
same architecture as vla_3f, trained jointly on three tasks (9 instructions). final = 34 epochs (20k steps); epNNNN = intermediate |
same as vla_3f |
vla_3f_t3/ |
same architecture as vla_3f, trained on the third task only (3 instructions). final = 98 epochs (20k steps); epNNNN = intermediate |
same as vla_3f |
Usage: see the accompanying code repository (model_exchange/dualsys/). Base model: lerobot/smolvla_base.
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support