Refactor pretraining loaders for self-contained LeRobot v3 datasets

README.md

@@ -88,6 +88,8 @@ All architectures are selected via `configs/model/<framework>.yaml` with `archit
 -->
 
 ## News
+- **[2026/10/25]** 🤗 We release all cleaned robot datasets used for OpenWAM pretraining: [AgiBotWorld-Beta](https://huggingface.co/datasets/OpenWAM/AgiBotWorld-Beta), [RoboCOIN](https://huggingface.co/datasets/wayrise/RoboCOIN), [InternData-A1](https://huggingface.co/datasets/OpenWAM/InternData-A1), and [DROID](https://huggingface.co/datasets/OpenWAM/Droid).
+- **[2026/10/10]** 🌟 OpenWAM reaches **1K GitHub stars**! Thank you for your support!
 - **[2026/09/09]** 📄 OpenWAM [Paper](https://arxiv.org/pdf/2609.07398) is released on arXiv.
 - **[2026/09/06]** 🤖 OpenWAM is integrated into [XPolicyLab](https://github.com/XPolicyLab/XPolicyLab).
 - **[2026/09/06]** 🤗 We release all pretrained and finetuned models on [huggingface](https://huggingface.co/OpenWAM).
@@ -355,6 +357,22 @@ Wan2.2-TI2V-5B video backbone, and the Mutual attention mask.
    reports the returned action dimension and latency. Stop the deployment
    process after the checks.
 
+### Four-Source Pretraining
+
+Use the four cleaned LeRobot v3 datasets with the shared mixture configuration:
+
+~~~bash
+bash scripts/train.sh dataloader=mixture
+~~~
+
+[`configs/dataloader/mixture.yaml`](configs/dataloader/mixture.yaml) enables
+AgiBotWorld-Beta, RoboCOIN, DROID, and InternData-A1 with proportional sampling
+by default. Their standalone configurations are in `configs/dataloader/` and
+point to `/mnt/data/wangyuran/pretrain_data_final`. Override each source's
+`dataset_dir` when moving the datasets, keeping the supplied `meta` statistics
+with them. Readers use the existing statistics directly and require no trim
+CSV, exclusion list, or statistics generation step.
+
 ### OpenWAM-α Fine-Tuning
 
 1. Download LIBERO as shown above.

configs/dataloader/agibotworld.yaml

@@ -0,0 +1,29 @@
+type: agibotworld
+
+# Family root with native LeRobot v3 buckets and meta/stats_g2a.json.
+# The reader retains separate action/state statistics and native hand/mobile masks.
+dataset_dir: /mnt/data/wangyuran/pretrain_data_final/AgiBotWorld-Beta
+
+# Windows preserve episode boundaries and align video with action/state samples.
+num_frames: 33
+video_stride: 4
+window_stride: 1
+height: 384
+width: 320
+multiview: true
+split: train
+
+# Stored normalization statistics are read from the dataset's meta directory.
+normalize_mode: quantile
+enable_action_supervision: true
+unify_action: true
+
+# One consistent augmentation per training clip.
+color_jitter:
+  brightness: 0.2
+  contrast: 0.2
+  saturation: 0.2
+  hue: 0.0
+
+# Optional reproducible footage budget; null loads the complete dataset.
+total_hours: null

configs/dataloader/interndata_a1.yaml

@@ -0,0 +1,30 @@
+type: interndata_a1
+
+# Family root with native LeRobot v3 buckets and meta/stats_<embodiment>.json.
+# Raw arm poses use *_ee_to_robot_pose, independently of the gripper motion.
+dataset_dir: /mnt/data/wangyuran/pretrain_data_final/InternData-A1
+
+# Windows preserve episode boundaries and align video with action/state samples.
+num_frames: 33
+video_stride: 4
+window_stride: 1
+height: 384
+width: 320
+multiview: true
+split: train
+
+# Stored normalization statistics are read from the dataset's meta directory.
+normalize_mode: quantile
+enable_action_supervision: true
+unify_action: true
+unify_action_map: ["0-9", "34-43"]
+
+# One consistent augmentation per training clip.
+color_jitter:
+  brightness: 0.2
+  contrast: 0.2
+  saturation: 0.2
+  hue: 0.0
+
+# Optional reproducible footage budget; null loads the complete dataset.
+total_hours: null

configs/dataloader/oxe_droid.yaml

@@ -0,0 +1,30 @@
+type: oxe_droid
+
+# Single LeRobot v3 dataset with meta/eef_stats.json and resolved tasks.parquet.
+# Wrist/mount poses and inverted closedness retain the canonical DROID EEF contract.
+dataset_dir: /mnt/data/wangyuran/pretrain_data_final/Droid
+
+# Windows preserve episode boundaries and align video with action/state samples.
+num_frames: 33
+video_stride: 4
+window_stride: 1
+height: 384
+width: 320
+multiview: true
+split: train
+
+# Stored normalization statistics are read from the dataset's meta directory.
+normalize_mode: quantile
+enable_action_supervision: true
+unify_action: true
+unify_action_map: ["0-9", "34-43"]
+
+# One consistent augmentation per training clip.
+color_jitter:
+  brightness: 0.2
+  contrast: 0.2
+  saturation: 0.2
+  hue: 0.0
+
+# Optional reproducible footage budget; null loads the complete dataset.
+total_hours: null