Shorten README to the final result

README.md

@@ -1,120 +1,45 @@
 ---
 license: cc-by-4.0
-pretty_name: FAR PRISM Data (PRISM-137, success-filtered)
+pretty_name: FAR PRISM Data
 task_categories:
   - robotics
 tags:
   - humanoid
   - real-to-sim
   - human-object-interaction
-  - video-to-video
   - counterfactual-video
   - unitree-g1
-  - motion-retargeting
-  - imitation-learning
 size_categories:
   - n<1K
 ---
 
 # FAR PRISM Data
 
-Training data for **PRISM**: counterfactual human–object interaction videos, their real-to-sim reconstructions, and the retargeted humanoid demonstrations used to train a privileged teacher policy and a depth-based student policy on a 29-DoF Unitree G1.
-
-**Release status:** only `data/train-student/` (student training data and reference replays) is published so far. `data/raw_video/`, `data/recon-hoi/` and `data/train-teacher/` are under review and currently hold placeholder READMEs.
-
-This release covers the **PRISM-137** batch: 137 counterfactual videos were generated from 5 real seed videos of a person carrying a box, by prompting a video-to-video model to swap the carried object for a **box, bin, barrel or ball**. All 137 clips were reconstructed and retargeted; **only the 129 clips whose teacher rollout succeeded are published here**, because failed rollouts were not used for student training.
+Training data for **PRISM**: counterfactual human–object interaction videos turned into humanoid (Unitree G1, 29 DoF) demonstrations for picking up, carrying and putting down boxes, bins, barrels and balls.
 
 <p align="center">
-  <img src="previews/montage_2x2.gif" alt="One example per category (box, bin, barrel, ball): counterfactual video on the left, reference replay of the student training motion on the right, time-aligned" width="960"/>
+  <img src="previews/montage_2x2.gif" alt="One example per category: counterfactual video (left) and reference replay of the student training motion (right)" width="960"/>
 </p>
 
-One example per category (box, bin, barrel, ball). Left: the counterfactual video; right: the reference replay of the student training motion, time-aligned to the video. Clip ids are listed in `previews/preview_clips.txt`; an MP4 version is in `previews/montage_2x2.mp4`.
-
-## Success filtering
-
-Each of the 137 clips was rolled out by the teacher checkpoint (`zihanw22/carry-any/ch2ckwzw`, `model_40000`) in simulation. The rollout is graded with two criteria:
-
-| Criterion | Definition | Count |
-| --- | --- | ---: |
-| `core_success` | stable robot–object contact **and** final object position within 0.5 m of the reference | **129 / 137** |
-| `strict_success` | `core_success` **and** no foot–object contact during the middle 20–80 % of the clip | 109 / 137 |
+Left: counterfactual video. Right: reference replay of the student training motion. One example per category; clip ids in `previews/preview_clips.txt`.
 
-The published subset is `core_success == true`. This matches the success definition used in the paper (complete pick–carry–drop and place the object at the target). The `strict_success` flag is kept as a column so you can filter further.
+## What is included
 
-| Category | Generated | Published (core) | Strict |
-| --- | ---: | ---: | ---: |
-| box | 35 | 35 | 31 |
-| bin | 34 | 32 | 26 |
-| barrel | 34 | 34 | 31 |
-| ball | 34 | 28 | 21 |
-| **total** | **137** | **129** | **109** |
-
-The 8 excluded clips (all `failed_no_stable_contact`) are listed in `excluded_clips.csv`. Per-clip status and object-position error live in `clips.csv`; aggregate numbers are in `success_statistics.json`.
-
-## Layout
+**129 clips** (box 35, bin 32, barrel 34, ball 28), all successful teacher rollouts. Per-clip index: `clips.csv`.
 
 ```text
-clips.csv                      one row per published clip: category, success flags, and every path below
-excluded_clips.csv             the 8 clips that failed the teacher rollout (not included)
-success_statistics.json        aggregate success statistics
-previews/                      montage_2x2.gif / .mp4 preview used above + preview_clips.txt
-data/
-  raw_video/                   UNDER REVIEW – placeholder README only (counterfactual videos + seed videos)
-  recon-hoi/                   UNDER REVIEW – placeholder README only (object meshes, URDFs, retarget results)
-  train-teacher/               UNDER REVIEW – placeholder README only (teacher motion bank)
-  train-student/
-    data/                      everything the student distillation stage consumed
-      motion_bank/             teacher rollouts of the 129 clips: <clip_id>.npz, 359 frames at 50 fps,
-                               plus precomputed joystick commands (policy_command_xy_yaw, policy_command_phase)
-      contact_sidecars/clips/<clip_id>/   contact intervals and contact points per body part, rollout metadata
-      robot_assets/            G1 URDF and depth-rendering meshes used by the student
-      checkpoints/             teacher_40000.pt (privileged teacher) and box_23000.pt (student initialiser)
-    vis/                       129 reference-replay videos of the exact student motion bank  <clip_id>.mp4  (1280x720, 359 frames, 50 fps)
+data/train-student/
+  data/motion_bank/          129 x <clip_id>.npz  (359 frames, 50 fps, robot + object trajectories, joystick commands)
+  data/contact_sidecars/     per-clip contact intervals and contact points
+  data/robot_assets/         G1 URDF and depth-rendering meshes
+  data/checkpoints/          teacher_40000.pt, box_23000.pt
+  vis/                       129 x <clip_id>.mp4  reference replays (1280x720, 50 fps)
+data/raw_video/  data/recon-hoi/  data/train-teacher/     under review, placeholder README only
 ```
 
-Clip ids follow `prism_cf_<category>_m<seed>_v<variant>`: `m` is the seed video index the counterfactual was generated from, `v` the generation variant.
-
-## Data formats
-
-**Motion NPZ (teacher and student banks).** Keys: `fps`, `joint_names` (29), `body_names` (32), `joint_pos (T,36)`, `joint_vel (T,35)`, `body_pos_w (T,32,3)`, `body_quat_w (T,32,4)`, `body_lin_vel_w`, `body_ang_vel_w`, `object_pos_w (T,3)`, `object_quat_w (T,4)`, `object_lin_vel_w`, `object_ang_vel_w`, `object_size (3,)`, `object_name`, `object_category`, `object_urdf_path`. Teacher clips have T = 360, student clips T = 359. Student clips additionally carry `policy_command_xy_yaw (T,3) float32` and `policy_command_phase (T,) uint8` (turn-then-forward command timeline).
-
-**Contact sidecars.** `contact_intervals.json` gives `[start, end]` step ranges per body part (`-1` = no contact); `*_contact_points.npy` / `*_contact_point_counts.npy` / `*_contact_interval_steps.npy` hold the object-frame contact points; `metadata.json` records the rollout grading (`success`, `stable_contact_success`, `final_position_success`, `no_middle_foot_object_contact_success`, position/rotation error, thresholds).
-
-**Object maps.** `_clip_object_urdf_map.json` in each motion bank resolves a clip to its URDF (`_single_slot_urdfs/`, included) and to its visual and collision meshes under `data/recon-hoi/<clip_id>/`, which is under review and not yet uploaded; mesh-dependent loading needs that directory in place. Object size, mass priors and inertia are already available from the object map and URDFs. Absolute paths that remain inside NPZ metadata, `contact_sidecars/*/metadata.json` and the `_source_*` / `source_manifest_137.json` files are historical provenance and are not needed for loading; the object map is authoritative.
-
-## Usage
+Training code: <https://github.com/Z1hanW/holosoma> (branch `release/4kocpixa-clean`).
 
 ```python
 from huggingface_hub import snapshot_download
 root = snapshot_download("ImNotPrepared/far-prism-data", repo_type="dataset")
-
-import csv, numpy as np
-clips = list(csv.DictReader(open(f"{root}/clips.csv")))
-strict = [c for c in clips if c["strict_success"] == "True"]
-d = np.load(f"{root}/{clips[0]['student_motion']}")
-print(d["joint_pos"].shape, d["object_pos_w"].shape, d["policy_command_xy_yaw"].shape)
-```
-
-Training code: <https://github.com/Z1hanW/holosoma>, branch `release/4kocpixa-clean`, commit `572c385ec85c94469c69851c762a0aee2e49608a`. Point `--motion-bank` at `data/train-student/data/motion_bank` with `--contact-bank data/train-student/data/contact_sidecars`, `--robot-assets data/train-student/data/robot_assets` and the two checkpoints (student). The original run used an 8-rank student shard manifest generated over all 137 clips; those manifests are not included because this release drops 8 clips, so regenerate rank shards for the 129-clip bank with the HoloSoma `prepare_as_rank_shards` tooling before launching multi-GPU training.
-
-## Provenance
-
-- Seed videos: 5 real recordings of a person picking up, carrying and putting down a box.
-- Counterfactual videos: video-to-video generation replacing the carried object with a box, bin, barrel or ball while preserving the scene, camera and motion.
-- Reconstruction: CRISP-style real-to-sim pipeline with contact anchors (human mesh recovery, SAM3D object initialisation, six-axis mask-IoU refinement, support-plane grounding, COACD collision meshes). Run `prism_axis_hybrid_object_humanscale_*_20260826`.
-- Retargeting: contact-anchored, all-frame-grounded retargeting to the Unitree G1 with exact non-penetration; HoloSoma bank `prism_world_zhen137_humanscale_w50_smooth5_20260826`.
-- Teacher: privileged robot–object tracking policy trained on the 137-clip bank (`zihanw22/carry-any/ch2ckwzw`).
-- Student data: native rollouts of `ch2ckwzw/model_40000` on the 137 clips with precomputed turn-then-forward commands and runtime contact sidecars; success grading as above.
-
-The reconstruction stage (object meshes, URDFs, retargeted robot–object trajectories), the counterfactual videos and the teacher bank are under review; the intermediate human-mesh reconstructions of the CRISP run are not released.
-
-## Citation
-
-```bibtex
-@inproceedings{prism2026,
-  title     = {PRISM: Scaling Humanoid Object Carrying from Counterfactual Videos},
-  author    = {Wang, Zihan and others},
-  booktitle = {Conference on Robot Learning (CoRL)},
-  year      = {2026}
-}
 ```

SHA256SUMS

@@ -1,4 +1,3 @@
-0595b61579065a12bd9bf9c372d8133cce78fd0b4a21b64444b803ad406d7272  README.md
 0364df82e5c79658c8c843a8b81b683bea0d66ccc754f9e11fc319f9a9eed1f9  clips.csv
 e9de2954556f7f39c98cc5e90de2e28550dad4ba656c986280918c929af1256d  data/train-student/data/checkpoints/box_23000.pt
 14e644323b8e6a7b769dbf641d9f625bee895742b7064368b55afd0710a0d665  data/train-student/data/checkpoints/teacher_40000.pt
@@ -4880,3 +4879,4 @@ c254dd86d20735b1916b4803436bb3833c8e1731802ec1a1262546a88784e8b3  success_statis
 4e659724f412ea1a63f0281c1cfcfa0f1693403af6ca1ea446ba367ce45fbd40  data/raw_video/README.md
 f84c48e6c175e90136652f98ead8da0b8757b233fcf282aca3fac758d3e19b23  data/recon-hoi/README.md
 5f2bd380568083b82a161c69fe1c5b9031d0ae46338b233196e8028d30cdba4f  data/train-teacher/README.md
+2ec177ad48bfc57a3ef5c1fadf191c0d682b9dfb44b247b793450030b23cdcd6  README.md