Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,30 @@
+---
+tags:
+- robotics
+- lerobot
+- reinforcement-learning
+---
+
+# Reversed-curiosity RL assets for LGES case pick (icm_0708.pt / critic_0708.pt)
+
+Companion checkpoints for RL fine-tuning of `Chanho-Lee/smolvla_naive_0708`
+(reversed-curiosity / anti-exploration reward, AWR + BC). Code:
+`smolvla_meanflow` plugin.
+
+| file | class | contents |
+|---|---|---|
+| `icm_0708.pt` | `smolvla_meanflow.icm.ICM` | frozen forward models, streams ['proprio', 'vision'], per-stream calibrated eta {'proprio': 1.6751, 'vision': 0.3907} |
+| `critic_0708.pt` | `smolvla_meanflow.critic.CriticEnsemble` | V(s) on the fixed reward mix {'proprio': 1.0, 'vision': 1.0} (input: concat features, 975-d) |
+
+Reward: `r = sum_s w_s * exp(-eta_s * err_s) / sum_s w_s` in (0, 1] per step.
+Trained on `Chanho-Lee/lges_case_pick_0708` with the last episode of each
+pallet layer held out ([14, 29, 44, 59, 74]).
+
+```python
+from huggingface_hub import hf_hub_download
+from smolvla_meanflow.icm import ICM
+from smolvla_meanflow.critic import CriticEnsemble
+
+icm = ICM.load(hf_hub_download("Chanho-Lee/lges_case_pick_0708_rl", "icm_0708.pt"))
+critic = CriticEnsemble.load(hf_hub_download("Chanho-Lee/lges_case_pick_0708_rl", "critic_0708.pt"))
+```