Update benchmark README

README.md

@@ -1,18 +1,81 @@
 ---
-pretty_name: Video-Progress Benchmark
+pretty_name: Video Progress Benchmark
 tags:
-- robotics
-- video-progress
-- benchmark
-- embodied-ai
+  - robotics
+  - embodied-ai
+  - video-progress
+  - progress-estimation
+  - reward-modeling
+  - benchmark
 license: other
 ---
 
-# Video-Progress Benchmark Release
+# Video Progress Benchmark (VPB)
 
-Video-Progress Benchmark is a benchmark for evaluating progress prediction in long-horizon robot manipulation. This release provides the benchmark split files, the minimal subset of raw videos required by the benchmark protocol, and a frame-extraction workflow for reproducible evaluation of VLAC-Cut benchmark metrics.
+**A video-language benchmark for process-level robot task progress estimation.**
 
-## Contents
+[Code](https://github.com/InternRobotics/VLAC-cut) · [Paper](<PAPER_URL>) · [VLAC-Cut Model](https://huggingface.co/InternRobotics/VLAC-Cut)
+
+The **Video Progress Benchmark (VPB)** evaluates whether a model can estimate how a robot task evolves throughout a video—not only whether the final frame looks successful. VPB is built from the held-out portion of the **Progress Annotation Dataset** and explicitly tests advancement, stagnation, regression, and recovery.
+
+This release provides the official split files, the raw-video subset required by the benchmark protocol, frame-extraction utilities, and scripts for reproducible VLAC-Cut inference and metric computation.
+
+## Why VPB?
+
+Most existing robot reward and progress evaluations rely on endpoint success, isolated observations, frame pairs, or trajectories that are assumed to improve monotonically with time. Real robot executions are often different: a missed grasp, collision, drop, or wrong-object interaction can leave progress unchanged or move the task backward before a later recovery.
+
+VPB addresses this gap through two design choices:
+
+1. **Continuous video context:** models evaluate temporally ordered robot videos rather than disconnected images.
+2. **Process-level evaluation:** metrics separately measure global trajectory recovery, terminal-state recognition, and local event-level direction.
+
+## Progress Annotation Dataset
+
+Each annotated record contains a task instruction, a task-level plan, and task-relevant progress points. A progress point is attached to an observable semantic event such as contact formation or loss, grasp or placement outcome, object displacement, failure, stagnation, regression, or recovery.
+
+When available, progress points also include auxiliary process supervision: scene state, robot action or object interaction, progress explanation, success/failure analysis, correction plan, and task-relevant grounding.
+
+## Dataset Scale
+
+| Inventory | Records | Episodes | Tasks | Keyframe progress points |
+|---|---:|---:|---:|---:|
+| Full annotation inventory | 35,230 | 26,615 | 15,206 | 464,446 |
+| Curated train/evaluation inventory | 28,167 | 22,978 | 15,206 | 375,172 |
+
+## Official Splits
+
+VPB is organized along two axes:
+
+- **Seen vs. unseen:** whether the semantic task unit appears in the training split.
+- **Expert vs. non-expert:** whether the annotated trajectory contains a regressive progress transition.
+
+| Split | Records | Episodes | Keyframe progress points |
+|---|---:|---:|---:|
+| Train | 24,652 | 20,479 | 331,762 |
+| Expert seen | 1,043 | 1,033 | 11,769 |
+| Expert unseen | 1,043 | 1,035 | 12,333 |
+| Non-expert seen | 713 | 706 | 9,458 |
+| Non-expert unseen | 716 | 706 | 9,850 |
+
+All views from the same physical execution are assigned to the same split. Held-out progress annotations are excluded from prompt construction, augmentation, in-context demonstration selection, fine-tuning, and checkpoint selection.
+
+## Evaluation Protocol
+
+Annotated semantic keyframes are converted into a canonical reference trajectory by piecewise-linear interpolation. This interpolation is an evaluation convention; it does not assume that physical progress changes linearly between events.
+
+- global and terminal metrics are computed on the official `1 Hz` evaluation grid;
+- local direction metrics are computed directly on adjacent annotated semantic anchors;
+- metrics are first computed per record and then macro-averaged, preventing long videos from dominating the result.
+
+### Metrics
+
+| Scope | Metrics | What they measure |
+|---|---|---|
+| Global trajectory | MAE, PRC, VOC | Absolute calibration and ordering of progress states; VOC is reported only for expert trajectories |
+| Terminal state | TSA, successful F1, failed/incomplete F1, Macro-F1 | Whether the final state is complete using the common `≥ 90` threshold |
+| Local direction | AP+, AP−, MacroAP | Whether adjacent semantic events are correctly ranked as improvement or regression |
+
+## Repository Contents
 
 ```text
 benchmark_splits/
@@ -39,12 +102,7 @@ data/
   test_videos.tar
 ```
 
-## Usage
-
-VLAC-Cut inference uses the model release's Transformers interface and requires
-an environment that can load `Qwen3VLMoeForConditionalGeneration`, plus enough
-GPU memory for the 30B checkpoint. Frame extraction and metric computation use
-only the files in this benchmark release.
+## Quick Start
 
 ### 1. Unpack the video archives
 
@@ -52,15 +110,13 @@ only the files in this benchmark release.
 bash scripts/unpack_data.sh
 ```
 
-This command extracts `data/train_videos.tar` and `data/test_videos.tar` into `data/`.
-
-To extract the raw videos elsewhere, pass a target directory:
+By default, the archives are extracted under `data/`. To use another location:
 
 ```bash
 bash scripts/unpack_data.sh /path/to/data
 ```
 
-### 2. Extract frames
+### 2. Extract benchmark frames
 
 ```bash
 python scripts/extract_vlac2_release_frames.py \
@@ -68,21 +124,27 @@ python scripts/extract_vlac2_release_frames.py \
   --frames-root /path/to/data_extracted_frames
 ```
 
-This command extracts the benchmark main-view frames from `/path/to/data` into `/path/to/data_extracted_frames`. If `--frames-root` is omitted, frames are written to `_extracted_frames/` under the release directory.
+The script extracts the benchmark main-view frames. If `--frames-root` is omitted, frames are written to `_extracted_frames/` under the release directory.
 
-### 3. Use the benchmark splits
+The canonical split JSON files use the path prefix:
 
-The benchmark JSON files are pre-generated and can be used directly.
+```text
+__VLAC2_FRAMES_ROOT__/
+```
 
-Frame paths in the JSON files use the prefix:
+Resolve this prefix to the absolute extracted-frame directory. For example:
 
 ```text
-__VLAC2_FRAMES_ROOT__/
+__VLAC2_FRAMES_ROOT__/episode/frame.jpg
 ```
 
-Replace `__VLAC2_FRAMES_ROOT__` with the absolute path to the extracted-frame directory. For example, `__VLAC2_FRAMES_ROOT__/...` should be resolved as `/path/to/data_extracted_frames/...`.
+becomes:
 
-### 4. Build a VLAC-Cut evaluation manifest
+```text
+/path/to/data_extracted_frames/episode/frame.jpg
+```
+
+### 3. Build the VLAC-Cut evaluation manifest
 
 ```bash
 python scripts/build_vlac_cut_eval_manifest.py \
@@ -91,15 +153,13 @@ python scripts/build_vlac_cut_eval_manifest.py \
   --out manifests/vlac_cut_vpb_eval.jsonl
 ```
 
-For our VLAC-Cut benchmark evaluation, the manifest samples 2Hz video input frames, records the public 1Hz evaluation frames, and materializes the exact `chunk_all` text prompt in `vlac_cut_prompt`. The 2Hz setting is an evaluation choice for VLAC-Cut input, not a benchmark-wide requirement. Metrics are computed on the released 1Hz evaluation frames.
+For the reported VLAC-Cut evaluation, the manifest samples model-input frames at `2 Hz`, stores the public `1 Hz` evaluation frames, and materializes the exact `chunk_all` prompt in `vlac_cut_prompt`.
 
-For trajectory-level VLAC-Cut predictions, keep the manifest's 2Hz `frames` list in each prediction row. The evaluator maps predictions by original frame id and scores only the public 1Hz `eval_frames` for global progress and terminal metrics.
+The `2 Hz` input rate is a VLAC-Cut inference choice, not a benchmark-wide requirement. Official metrics are computed on the released `1 Hz` evaluation grid.
 
-By default, each manifest row is one trajectory with a list of sampled frame paths for VLAC-Cut inference.
+### 4. Run VLAC-Cut batch inference
 
-### 5. Run VLAC-Cut batch inference
-
-VLAC-Cut is released separately at <https://huggingface.co/InternRobotics/VLAC-Cut>.
+Download or clone the model release from `InternRobotics/VLAC-Cut`, then run:
 
 ```bash
 python scripts/run_vlac_cut_batch.py \
@@ -108,13 +168,19 @@ python scripts/run_vlac_cut_batch.py \
   --out predictions/vlac_cut_predictions.jsonl
 ```
 
-The batch prediction file uses one fixed schema:
+Each prediction row uses the following schema:
 
 ```json
-{"global_episode_id": "...", "frames": [65, 80, 95], "response": "时间: 0.0s, 进度: 0%\n时间: 1.0s, 进度: 50%"}
+{
+  "global_episode_id": "...",
+  "frames": [65, 80, 95],
+  "response": "时间: 0.0s, 进度: 0%\n时间: 1.0s, 进度: 50%"
+}
 ```
 
-### 6. Evaluate VLAC-Cut predictions
+Keep the manifest's original `frames` list in each trajectory-level prediction. The evaluator maps predictions by original frame ID and scores only the official `eval_frames`.
+
+### 5. Evaluate predictions
 
 ```bash
 python scripts/evaluate_vpb_predictions.py \
@@ -124,11 +190,30 @@ python scripts/evaluate_vpb_predictions.py \
   --out-md reports/vlac_cut_vpb_eval.md
 ```
 
-The evaluator reports global progress metrics, terminal success metrics, and local direction AP on adjacent semantic anchors. Global progress is shown for the 4-bucket overall split and each bucket; terminal metrics and local direction AP are shown for 4-bucket overall, seen merged, and unseen merged. See `docs/evaluate_vlac_cut_on_vpb.md` for the accepted prediction schema and metric definitions.
+The evaluator reports:
+
+- global progress metrics for the four-bucket overall split and each individual bucket;
+- terminal metrics for four-bucket overall, seen merged, and unseen merged;
+- local direction AP for four-bucket overall, seen merged, and unseen merged.
+
+See `docs/evaluate_vlac_cut_on_vpb.md` for accepted prediction formats and complete metric definitions.
+
+## Reference VLAC-Cut Results
+
+| Global MAE ↓ | Global PRC ↑ | Terminal Macro-F1 ↑ | Local Direction MacroAP ↑ |
+|---:|---:|---:|---:|
+| **7.3344** | **0.9230** | **73.75** | **71.01** |
+
+These values summarize the overall VPB evaluation. Terminal Macro-F1 and Local Direction MacroAP are reported in percent. Refer to the paper for per-bucket results and baseline comparisons.
+
+## Citation
+
+Please replace the placeholder below with the final paper BibTeX before public release.
+
+```bibtex
+
+```
 
-## Notes
+## License
 
-* This repository contains only the raw videos required by the released benchmark splits.
-* The full raw-data release is distributed separately.
-* The benchmark JSON files are canonical release artifacts and do not need to be rebuilt.
-* The raw-video archives and the extracted-frame cache are intentionally separated so that users can regenerate frames under their own storage layout.
+See the license files distributed with this release. Third-party datasets and videos retain their original licenses and usage restrictions.