LINUAN/fold_towels_recap_value_scalar

fold_towels_recap value network — SCALAR head (A/B variant)

Scalar-head value network for the SO-101 fold-towels RECAP (pi-star 0.6) reproduction. This is the 1-dim MLP (MSE regression) variant, NOT the 201-bin distributional head.

• Model class: RECAPValueNetworkScalar (see recap_value_network_scalar.py)
• Config class: RECAPValueNetworkConfig (note: num_value_bins=201 is inherited from the distributional config for identical hyperparameters, but the actual head is 1-dim)
• Backbone: PaliGemma gemma_2b, vision encoder + backbone frozen, 3 backbone layers unfrozen
• Data: LINUAN/fold_towels_recap_corpus_v2 (276 episodes, val split 0.1)
• Trained 3 epochs; final val_mae=0.05776, val_loss(MSE)=0.00587

Load

from lerobot.rl.algorithms.recap_value_network_scalar import RECAPValueNetworkScalar
from lerobot.rl.algorithms.recap_value_network import RECAPValueNetworkConfig
import torch, json
cfg = RECAPValueNetworkConfig(**json.load(open("config.json")))
model = RECAPValueNetworkScalar(cfg)
state = torch.load("model.safetensors")  # or use safetensors.torch.load_file
model.load_state_dict(state)