Add project-page summary to the dataset card

README.md

@@ -2,18 +2,57 @@
 pretty_name: LoopWAM checkpoints
 ---
 
-# LoopWAM checkpoints
+# LoopWAM
 
-Released checkpoints for [LoopWAM](https://loopwam.github.io/). Training code: [loop-wam/LoopWAM](https://github.com/loop-wam/LoopWAM).
+**Closed-loop generalist policy improvement from deployment rollouts.**
 
-Each directory contains one `.pt` weight and its `dataset_stats.json`.
+[Project page](https://loopwam.github.io/) · [Code](https://github.com/loop-wam/LoopWAM)
+
+LoopWAM turns deployment into continued learning. Expert demonstrations, together with successful, suboptimal, and failed rollouts, train one world–action model. Every rollout supervises future video. Action imitation is weighted by trajectory quality, and failed actions are not imitated.
+
+Demos, task videos, and paired action–video generation are on the [project page](https://loopwam.github.io/).
+
+## Results
+
+Best overall performance among the methods compared in the paper.
+
+| Benchmark | LoopWAM |
+| --- | --- |
+| [RoboChallenge Table30-V2](https://loopwam.github.io/) | **43.33%** success · **56.80** score |
+| [RoboTwin 2.0](https://loopwam.github.io/) | **94.5%** average (clean 94.6, randomized 94.4) |
+| [RoboCasa365](https://loopwam.github.io/) | **49.5%** overall |
+
+### Table30-V2 self-improvement
+
+30 real tasks, four embodiments, one multi-task model per embodiment. Three deployment rounds:
+
+| Stage | Success | Score |
+| --- | ---: | ---: |
+| Expert-trained | 23.67% | 36.10 |
+| Round 1 | 28.00% | 42.63 |
+| Round 2 | 37.67% | 51.65 |
+| Round 3 | 43.33% | 56.80 |
+
+Per-embodiment success at Round 3: ALOHA 26.00, DOS-W1 42.00, ARX5 55.71, UR5 76.67. Overall score uses task weights 10 / 10 / 7 / 3.
+
+Shirt folding, after single-task post-training, rises from 13.20% to 32.54% and then to **55.56%** across two further deployment rounds.
+
+## Method
+
+Video and action tokens interact through bidirectional attention while they are denoised together.
+
+- **Source-dependent noise.** Expert data trains joint generation. Lower-noise rollout actions teach the world model the consequences that were observed.
+- **Quality-aware learning.** Quality conditions both streams. Action loss increases with quality. Failed rollouts keep full video supervision and receive no direct action-imitation loss.
+- **Expert-quality steering.** At deployment, an expert-quality target steers the policy.
+
+## Checkpoints
+
+Each directory contains one weight file and its `dataset_stats.json`.
 
 | Benchmark | Directory | Checkpoint |
 | --- | --- | --- |
-| RoboChallenge Table30-V2, UR5 | `table30v2_ur5/` | `step_029925.pt` |
-| RoboChallenge Table30-V2, ALOHA | `table30v2_aloha/` | `step_061320.pt` |
-| RoboChallenge Table30-V2, ARX5 | `table30v2_arx5/` | `step_057355.pt` |
-| RoboChallenge Table30-V2, DOS-W1 | `table30v2_w1/` | `step_072970.pt` |
+| Table30-V2, UR5 | `table30v2_ur5/` | `step_029925.pt` |
+| Table30-V2, ALOHA | `table30v2_aloha/` | `step_061320.pt` |
+| Table30-V2, ARX5 | `table30v2_arx5/` | `step_057355.pt` |
+| Table30-V2, DOS-W1 | `table30v2_w1/` | `step_072970.pt` |
 | RoboTwin 2.0 | `robotwin/` | `step_037645.pt` |
-
-Table30-V2: 43.33% success, score 56.80. RoboTwin 2.0: 94.6% clean, 94.4% randomized, 94.5% average.