Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,31 @@
+---
+license: mit
+tags: [diffusion-policy, robotics, franka]
+---
+# oopsie-debug
+
+Diffusion Policy (image, UNet) checkpoints trained on the Franka teleop block dataset
+`teleop_dataset_20260920_205906` (20 episodes, 1206 steps @ 10 Hz, all demos used for training,
+no held-out validation). Trained 2001 epochs (~34k steps, batch 64); loss values in the table are
+in-sample train loss.
+
+| file | epoch | train loss |
+|---|---|---|
+| epoch_0500.ckpt | 500 | 0.0072 |
+| epoch_1000.ckpt | 1000 | 0.0031 |
+| epoch_1500.ckpt | 1500 | 0.0011 |
+| epoch_2000.ckpt | 2000 | 0.0004 |
+
+Each `.ckpt` is a full training checkpoint (`state_dicts`: model, ema_model, optimizer;
+`pickles`: epoch, global_step, rng_state; `cfg`), loadable with
+`torch.load(path, pickle_module=dill, map_location='cpu', weights_only=False)`.
+Use `state_dicts['ema_model']` for inference.
+
+## Observation / action contract
+- obs: `img_third`, `img_wrist` (3x120x160 RGB, resized from 480x640 BGR source), `robot_eef_pose`
+  (9-D: xyz + rotation_6d), `gripper_state` (1-D width in m). n_obs_steps = 2.
+- action (10-D): `[x, y, z, rot6d(6), gripper]` = pose/gripper of frame t+1. horizon 16, 8 executed.
+- Normalization: position min/max -> [-1,1]; rotation_6d identity (no min/max);
+  gripper fixed [0, 0.08] m -> [-1,1]. The source quaternion is scalar-last (xyzw).
+- Code: `memory_diffusion_policy/dataset/teleop_block_image_dataset.py`
+  (`action10_to_pose7_gripper` is the inverse for rollout). Full config in `config.yaml`.