Robotics
Transformers
Safetensors
English
Chinese
vla
action-expert
vlabench
robocasa
cot
world-model
grpo
Instructions to use yixue888/Coract_checkpoint with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use yixue888/Coract_checkpoint with Transformers:
# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("yixue888/Coract_checkpoint", device_map="auto") - Notebooks
- Google Colab
- Kaggle
audit: with_state_configs/README.md
Browse files- with_state_configs/README.md +16 -0
with_state_configs/README.md
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# with_state_configs:6 个遗留 waypoint-regression run 的 epoch-10(last)checkpoint
|
| 2 |
+
|
| 3 |
+
各 run 的 SFT 配置在 GitHub [CoRAct_model](https://github.com/onlyxingxingzyx/CoRAct_model) `configs/with_state_configs/`。此目录为早期消融存档,每个 run 只上传了 epoch-10(last,step 点位均早于 ep10)的必需 4 文件。
|
| 4 |
+
|
| 5 |
+
## 成功率(copy_test 协议,17 任务 ~510 episodes,CoT+GT waypoint 回归)
|
| 6 |
+
|
| 7 |
+
| run | in-dist 成功率 | 备注 |
|
| 8 |
+
|---|---|---|
|
| 9 |
+
| `robocasa_full_attention_cot_xyz_waypoint_regression_state_text` | **81.6%** (416/510) | 无 world model(state_text)版,本地最高 |
|
| 10 |
+
| `robocasa_full_attention_cot_xyz_waypoint_regression_wm_state_text` | 80.2% (409/510) | = `robocasa/full_attention/with_state_configs/` 下同一 run(那里有 ep7/8/9/10 完整对比与 OOD 套件) |
|
| 11 |
+
| `robocasa_full_attention_cot_xyz_rotation_waypoint_regression_wm_state_text` | 72.4% (369/510) | xyz+rotation CoT 版(GRPO 实验的 SFT 基座家族) |
|
| 12 |
+
| `robocasa_full_attention_cot_xyz_future_left_waypoint_regression_state_text` | —(未单独评测) | future-left 目标表示消融 |
|
| 13 |
+
| `robocasa_full_attention_cot_xyz_future_left_waypoint_regression_wm_state_text` | —(未单独评测) | 同上 + world model |
|
| 14 |
+
| `vlabench_full_attention_cot_xyz_future_second_waypoint_regression_wm_state_text` | —(VLABench run,未用此协议评测) | 见 `vlabench/` |
|
| 15 |
+
|
| 16 |
+
评测口径与 `robocasa/README.md` 的 copy_test 相同(每任务 30 episodes)。OOD 评测只对 `xyz_waypoint_regression_wm_state_text` 的 epoch-8 做过(见 `robocasa/`)。
|