README.md

---
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"))
```