act-sotac-tac-mlp

ACT policy with one added transformer encoder token from a tactile encoder. Variant: MLP encoder on the current tactile frame (312 values), trained end to end.

Trained on episodes 0 to 70 of Jingyi-Z/sotac (red foam ball into container, SO-101, top and wrist cameras, Paxini fingertip forces at 30 Hz), 100k steps, batch 8, stock ACT hyperparameters, lerobot 0.4.4 with a tactile-encoder patch. The config carries tactile_key, tactile_hidden and related fields: apply the matching ACT patch before loading, or strip those keys for stock lerobot. Companion variants: act-sotac-tac-{mlp,pred,esn,esnpred}. Baseline without tactile: act-sotac-runA.

Downloads last month
4
Safetensors
Model size
52.4M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support