rewrite dataset card (no emojis, prose-style) + composite hero/grid assets

README.md

@@ -24,241 +24,134 @@ tags:
 - nvidia
 ---
 
-# 🏭 PhysicalAI SDG-Warehouse
-
-**A 412-hour, ~123K-clip synthetic warehouse safety video dataset — multi-view CCTV + eye-level rigs, four staged industrial-safety scenarios, full reproducibility from a single seed, packaged as streamable WebDataset shards.**
-
-<p align="center">
-  <img src="./assets/scenario_nearmiss.jpg"     width="24%" alt="Forklift–human near-miss"/>
-  <img src="./assets/scenario_fire.jpg"         width="24%" alt="Warehouse fire"/>
-  <img src="./assets/scenario_collision.jpg"    width="24%" alt="Forklift–shelf collision"/>
-  <img src="./assets/scenario_box_pickup.jpg"   width="24%" alt="Warehouse box pickup"/>
-</p>
-
-<p align="center">
-  <em>Forklift–human near-miss &nbsp;·&nbsp; Warehouse fire &nbsp;·&nbsp; Forklift–shelf collision &nbsp;·&nbsp; Box pickup</em>
-</p>
-
-<p align="center">
-  <a href="https://huggingface.co/datasets/nvidia/PhysicalAI-SDG-WareHouse"><img src="https://img.shields.io/badge/🤗%20HuggingFace-Dataset-yellow"></a>
-  <img src="https://img.shields.io/badge/license-CC--BY--4.0-blue">
-  <img src="https://img.shields.io/badge/clips-~123K-success">
-  <img src="https://img.shields.io/badge/duration-~412h-success">
-  <img src="https://img.shields.io/badge/resolution-1920×1080-success">
-  <img src="https://img.shields.io/badge/fps-30-success">
-  <img src="https://img.shields.io/badge/format-WebDataset-orange">
-  <img src="https://img.shields.io/badge/built%20with-NVIDIA%20Isaac%20Sim-76B900">
-</p>
+# PhysicalAI SDG-Warehouse
 
----
+![Hero — four warehouse-safety scenarios in synchronized multi-view simulation.](./assets/hero_2x2.gif)
 
-## At a glance
+PhysicalAI SDG-Warehouse is a synthetic, fully-annotated video dataset of staged industrial-safety events captured in a simulated warehouse environment. It contains approximately 123 thousand video clips, totaling roughly 412 hours of footage at 1920 by 1080 resolution and 30 frames per second, organized across four scenarios: a forklift near-miss with a human worker, a warehouse fire with worker evacuation, a forklift collision with a populated storage shelf, and a routine box-pickup action. Every multi-camera simulation run is filmed from five to ten synchronized viewpoints, and the entire pipeline is reproducible end-to-end from a single random seed.
 
-| | | | |
-|:--:|:--:|:--:|:--:|
-| **≈123K**<br>video clips | **≈412 h**<br>of footage | **≈29.2K**<br>runs (samples) | **4**<br>scenarios |
-| **1920×1080**<br>30 fps H.264 | **WebDataset**<br>tar shards (~5 GiB) | **5–10**<br>synced cameras / run | **100% reproducible**<br>(seed → pixels) |
+This dataset is described in Appendix A.1.5 of the Cosmos3 technical report (citation [below](#citation)).
 
-> **v1 release scope (this card):** RGB video only, organized as run-grouped WebDataset shards. The full **artifacts** tier — metric depth, instance + shaded segmentation, Canny edges, 2D/3D bounding boxes, and per-frame camera intrinsics+extrinsics — is rendered, ground-truth-aligned to every RGB frame, and will land in an `artifacts/` tier of the same repo as a follow-up. The annotation visualizations below are from those native renders.
+## Overview
 
----
+The release is packaged as standard [WebDataset](https://github.com/webdataset/webdataset) tar shards, with one sample per simulation run. Inside each shard, all of a run's synchronized camera views share the same sample key, so a single iteration of the dataset yields a complete multi-view group together with its run-level metadata. The shards are sized at approximately five gigabytes each, which is optimized for streaming directly into a training loop without first materializing the full dataset on disk.
 
-## ⚡ Quickstart
+This first release publishes the RGB video tier. A companion artifacts tier containing the full annotation stack — metric depth, instance and shaded segmentation, Canny edges, two-dimensional and three-dimensional bounding boxes, and per-frame camera intrinsics and extrinsics — is rendered natively by Isaac Sim and pixel-aligned to every RGB frame; it will land in an `artifacts/` directory of this same repository as a follow-up. The annotation visualizations later on this page show those native renders.
 
-Download a single scenario (a few hundred GB) — recommended for most users:
+The RGB tier consists of 459 WebDataset shards totaling approximately 2.24 tebibytes of video, plus two small Parquet indexes (one row per run and one row per camera-clip) that enable filtering by scenario, seed, camera, or source without opening a single shard.
 
-```bash
-pip install -U "huggingface_hub[hf_xet]"
+## Why this dataset
 
-huggingface-cli download nvidia/PhysicalAI-SDG-WareHouse \
-  --repo-type dataset \
-  --include "rgb/forklift_human_nearmiss/**" "metadata/**" \
-  --local-dir ./PhysicalAI-SDG-WareHouse
-```
+Real footage of warehouse safety incidents is operationally rare, logistically difficult to capture at scale, and legally sensitive to redistribute. Even when such footage exists, it is almost never paired with the kind of dense, per-pixel ground truth that physical-AI training pipelines benefit from: depth, instance identity, segmentation, edges, and tracked bounding boxes for every visible agent and prop. Building a real-world dataset that covers all of these signals across a balanced mix of incident types is, in practice, infeasible.
 
-Available scenario directories:
+We therefore generate the dataset in simulation. In a simulator, the event is guaranteed to happen, every event parameter is exposed and controllable, every camera viewpoint is precisely registered, and every frame is automatically paired with deterministic ground truth. Domain randomization over lighting, materials, asset choice, agent identity, agent motion, camera pose, and event parameters provides the variability that real-world capture would otherwise contribute. This approach lets us cover rare or operationally dangerous events — near-misses, evacuations, and rigid-body collisions — at scale, while keeping the dataset reproducible and extensible.
 
-| Scenario | Repo path | Shards | ~Size |
-|---|---|---:|---:|
-| Forklift–human near-miss | `rgb/forklift_human_nearmiss/` | 113 | ~549 GiB |
-| Warehouse fire | `rgb/warehouse_fire/` | 125 | ~619 GiB |
-| Forklift–shelf collision | `rgb/forklift_shelf_collision/` | 114 | ~559 GiB |
-| Warehouse box pickup | `rgb/warehouse_box_pickup/` | 107 | ~520 GiB |
+## Quickstart
 
-Or pull the whole dataset (~2.24 TiB):
+Download a single scenario, which is the recommended starting point for most users since each scenario is a few hundred gigabytes:
 
 ```bash
+pip install -U "huggingface_hub[hf_xet]"
+
 huggingface-cli download nvidia/PhysicalAI-SDG-WareHouse \
   --repo-type dataset \
+  --include "rgb/forklift_human_nearmiss/**" "metadata/**" \
   --local-dir ./PhysicalAI-SDG-WareHouse
 ```
 
-For **streaming** (no full download) see [Loading examples](#loading-examples).
-
----
-
-## Why this dataset?
-
-> Real industrial-safety footage is **rare, sensitive, and rarely annotated**. SDG-Warehouse fixes that by guaranteeing the event happens, controlling every parameter, and shipping per-pixel ground truth for every frame.
-
-- **🎯 Deterministic ground truth.** *(Coming with the artifacts tier.)* Depth, instance segmentation, shaded segmentation, Canny edges, and 2D / 3D bounding boxes — *natively rendered*, not predicted, perfectly aligned to the RGB.
-- **📹 Multi-view by construction.** Every run is filmed from 5–10 synchronized cameras (ceiling CCTV + worker-height eye-level / circular impact rigs), with stable per-camera aliases that match across releases.
-- **🔁 Reproducible.** Each run is driven by a single random seed that controls scene composition, lighting, agent identity, motion, camera pose, and event timing. Seed is exposed in every sample's `meta.json`.
-- **🚨 Rare-event coverage.** Forklift–human near-misses, fire evacuations, and shelf collisions — events that are operationally dangerous and legally fraught to capture in the real world.
-- **📦 Streamable.** Shipped as standard WebDataset `.tar` shards (~5 GiB each), so you can train without downloading hundreds of GB up front. A Parquet `metadata/clips.parquet` index lets you filter by scenario / camera / seed without opening a single tar.
-- **🏗️ Built on NVIDIA Isaac Sim** with Isaac Sim Replicator Object (IRO) and Replicator Agent (IRA), so the same pipeline can be extended to your own scenarios.
-
-This dataset is described in §A.1.5 of the [Cosmos3 technical report](#citation). <!-- TODO: replace #citation with arXiv URL when public -->
-
----
-
-## What you can build with it
-
-| Use case | Why SDG-Warehouse helps |
-|---|---|
-| **Industrial-safety event detection** (near-miss, collision, fire, evacuation) | Rare in real footage, fully covered here, with seed-controlled difficulty (e.g. last-moment dodge distance). |
-| **Multi-camera 3D tracking** | 5–10 synchronized camera viewpoints per run; full intrinsics + extrinsics ship with the artifacts tier. |
-| **Video / world foundation models (Cosmos and friends)** | 412 hours of physically-grounded, long-horizon, multi-camera video. |
-| **Forklift behavior / collision prediction** | Parameterized forklift trajectories, dodge events, and rigid-body shelf-impact dynamics. |
-| **Monocular depth, segmentation & edge models** | Pixel-perfect colorized depth, instance / shaded segmentation, and Canny edges — *available in the upcoming artifacts tier*, large-scale, perfectly aligned, automatic. |
-| **Sim-to-real & domain-randomization research** | Full per-light randomization (color temperature, intensity, exposure, color), per-asset variation, and parametric camera placement give a controlled testbed. |
-| **Synthetic-data ablations & scaling laws** | Independent reproducible runs make it straightforward to subsample, re-render, or augment for data-scaling studies. |
-| **VLM / video-LLM training & eval** | Dense scenario / camera / seed metadata in `metadata/clips.parquet`; per-frame captions can be derived from the run-level metadata. |
-
----
-
-## Modalities
-
-### Available now (this release — `rgb/` tier)
+The available scenario directories are summarized below.
 
-| Stream | Format | What it is |
-|---|---|---|
-| **RGB** | MP4 (H.264) | Photoreal Isaac Sim render of the scene, one mp4 per camera per run |
-| **Run metadata** | JSON (`meta.json`) | scenario, seed, source_kind, list of cameras with aliases + source paths, S3 provenance |
-| **IRA writer metadata** | TXT (`metadata.txt`) | Replicator-writer name / version / global seed (present for most multi-cam runs) |
-| **Per-clip Parquet index** | `metadata/clips.parquet` | one row per `(run × camera)` for fast filtering |
-| **Per-run Parquet index** | `metadata/runs.parquet` | one row per run with `clip_keys[]`, `n_cameras`, `total_bytes`, `shard_path_in_repo` |
+| Scenario | Repository path | Shards | Approximate size |
+|---|---|---:|---:|
+| Forklift–human near-miss | `rgb/forklift_human_nearmiss/` | 113 | 549 GiB |
+| Warehouse fire | `rgb/warehouse_fire/` | 125 | 619 GiB |
+| Forklift–shelf collision | `rgb/forklift_shelf_collision/` | 114 | 559 GiB |
+| Warehouse box pickup | `rgb/warehouse_box_pickup/` | 107 | 520 GiB |
 
-### Coming with the artifacts tier (`artifacts/<scenario>/*.tar`)
+To pull the full dataset (approximately 2.24 tebibytes), omit the `--include` filter. For streaming pipelines that never materialize the data on disk, see [Loading examples](#loading-examples).
 
-| Stream | Format (per camera per run) | What it is |
-|---|---|---|
-| **Metric depth** | `depth.mp4` (log-normalized colorized) | Per-pixel distance to camera |
-| **Instance segmentation** | `segmentation.mp4` (colorized) + `instance_seg.tar` (per-frame 16-bit PNG IDs) | Stable instance IDs across frames and views |
-| **Shaded segmentation** | `shaded_seg.mp4` | 3D-aware segmentation with normal-based shading |
-| **Canny edges** | `edges.mp4` | Canny edges computed on the shaded segmentation |
-| **2D / 3D bounding boxes** | `object_detection.json` (consolidated per camera) | Tight + loose 2D AABBs + oriented 3D OBBs, per agent / prop, per frame |
-| **Camera parameters** | `camera_params.json` (consolidated per camera) | Intrinsics + view transform, per frame |
+## Scenarios
 
-> **Note:** the README and `metadata/runs.parquet` schemas are forward-compatible — when the artifacts tier lands, the `run_id` and `clip_key` keys join cleanly to the RGB tier.
+Each scenario stages a different self-contained event inside a shared warehouse environment with shelves and props. Scenarios differ in which agents are spawned, what event is staged, and how the cameras are placed.
 
----
+### Forklift–human near-miss
 
-## Scenarios
+A worker stands at a fixed location while a forklift navigates along a planned path toward the same location. A configurable last-moment dodge distance distinguishes a near-miss from a direct-contact event, so the same scene composition can produce both event classes by varying a single parameter. Each multi-camera run is captured by a mixture of ceiling-mounted CCTV-style cameras (camera aliases `ceiling_00` through `ceiling_04`) and worker-height eye-level cameras (`eye_00` through `eye_04`).
 
-Each scenario is a self-contained event in a shared warehouse environment with shelves and props; scenarios differ in which agents are spawned, what event is staged, and how cameras are placed.
+### Warehouse fire
 
-### 🚜 Forklift–human near-miss
-A worker stands at a fixed location while a forklift navigates toward them along a planned path. A configurable **last-moment dodge distance** distinguishes a near-miss from an outright contact event, letting the same scene composition produce both event classes. Captured by a mix of ceiling-mounted CCTV-style cameras (`ceiling_00`–`ceiling_04`) and worker-height eye-level cameras (`eye_00`–`eye_04`).
+A localized volumetric fire ignites at a randomized position and time while a small crew of workers performs random walks. On ignition, each worker reacts: it orients toward the flame and then runs toward a designated exit waypoint along a navigation-mesh path. The result is rare emergency-response footage that combines dynamic flames, smoke, and coordinated human evacuation in a single shot. Cameras are placed at ceiling height to maximize floor coverage, with aliases `ceiling_00` through `ceiling_04`.
 
-### 🔥 Warehouse fire
-A localized volumetric fire ignites at a randomized position and time while a small crew of workers performs random walks. On ignition, each worker reacts — orienting toward the flame and then running toward a designated exit waypoint along a navigation-mesh path — producing rare emergency-response footage that combines dynamic flames, smoke, and coordinated human evacuation in a single shot. Cameras at ceiling height for floor coverage (`ceiling_00`–`ceiling_04`).
+### Forklift–shelf collision
 
-### 💥 Forklift–shelf collision
-A forklift drives at a parameterized initial distance toward a populated storage shelf and impacts it, producing visible **rigid-body knock-over and debris dynamics**. An optional character can be placed along the forklift's path to extend the scenario to a forklift–shelf–human three-body event. Cameras placed circularly around the impact site at varying heights, distances, and look-down angles (`cam_00`–`cam_05`).
+A forklift drives at a parameterized initial distance toward a populated storage shelf and impacts it, producing visible rigid-body knock-over and debris dynamics. An optional character can be placed along the forklift's path to extend the scenario to a three-body forklift–shelf–human event. Cameras are placed circularly around the impact site at varying heights, distances, and look-down angles, with aliases `cam_00` through `cam_05`.
 
-### 📦 Warehouse box pickup
-A worker navigates to a randomly placed box, performs a contact-rich pickup motion, and carries it through the warehouse. Provides routine, **non-incident action coverage** as a counterpoint to the three safety scenarios. Cameras (`cam_00`–`cam_09`) form a mixed CCTV + eye-level rig.
+### Warehouse box pickup
 
----
+A worker navigates to a randomly placed box, performs a contact-rich pickup motion, and carries the box through the warehouse. This scenario provides routine, non-incident action coverage as a counterpoint to the three safety scenarios. The camera rig is a mixed CCTV and eye-level configuration, with aliases `cam_00` through `cam_09`.
 
 ## Multi-view coverage
 
-Every run is captured by multiple synchronized cameras. Below: one forklift–human near-miss run from **5 ceiling-mounted CCTV** viewpoints (top) and **5 worker-height eye-level** viewpoints (bottom).
-
-<p align="center">
-  <img src="./assets/multiview_cctv_view1.jpg" width="18%"/>
-  <img src="./assets/multiview_cctv_view2.jpg" width="18%"/>
-  <img src="./assets/multiview_cctv_view3.jpg" width="18%"/>
-  <img src="./assets/multiview_cctv_view4.jpg" width="18%"/>
-  <img src="./assets/multiview_cctv_view5.jpg" width="18%"/>
-</p>
-<p align="center">
-  <img src="./assets/multiview_eyelevel_view1.jpg" width="18%"/>
-  <img src="./assets/multiview_eyelevel_view2.jpg" width="18%"/>
-  <img src="./assets/multiview_eyelevel_view3.jpg" width="18%"/>
-  <img src="./assets/multiview_eyelevel_view4.jpg" width="18%"/>
-  <img src="./assets/multiview_eyelevel_view5.jpg" width="18%"/>
-</p>
+Every multi-camera simulation run is captured from five to ten synchronized cameras. For the near-miss scenario, the rig consists of five ceiling-mounted CCTV-style cameras and five worker-height eye-level cameras, all pointed at the interaction. The figure below shows a single near-miss run from each of the ten viewpoints; because all cameras share a clock and the same scene, the same instant in time appears across all ten frames.
 
----
+![Multi-view coverage — one near-miss run from ten synchronized cameras: five ceiling and five eye-level.](./assets/multiview_grid.png)
 
-## Ground-truth modalities (per scenario)
+For the fire scenario, the rig is the five ceiling cameras only. For the forklift–shelf collision, six cameras are arranged circularly around the impact site at varying heights and look-down angles. For the box-pickup scenario, the rig is a mixed CCTV plus eye-level configuration with ten cameras.
 
-These previews show the **native renders** that ship in the upcoming artifacts tier; the RGB column is included with the current release.
+## Ground-truth modalities
 
-|              | RGB                                       | Depth                                  | Instance seg.                                    | Shaded seg.                                     | Canny edges                                |
-|--------------|-------------------------------------------|----------------------------------------|--------------------------------------------------|-------------------------------------------------|--------------------------------------------|
-| Near-miss    | <img src="./assets/scenario_nearmiss.jpg"   width="160"/> | <img src="./assets/nearmiss_depth.jpg"   width="160"/> | <img src="./assets/nearmiss_segmentation.jpg"   width="160"/> | <img src="./assets/nearmiss_shaded_seg.jpg"   width="160"/> | <img src="./assets/nearmiss_edges.jpg"   width="160"/> |
-| Fire         | <img src="./assets/scenario_fire.jpg"       width="160"/> | <img src="./assets/fire_depth.jpg"       width="160"/> | <img src="./assets/fire_segmentation.jpg"       width="160"/> | <img src="./assets/fire_shaded_seg.jpg"       width="160"/> | <img src="./assets/fire_edges.jpg"       width="160"/> |
-| Collision    | <img src="./assets/scenario_collision.jpg"  width="160"/> | <img src="./assets/collision_depth.jpg"  width="160"/> | <img src="./assets/collision_segmentation.jpg"  width="160"/> | <img src="./assets/collision_shaded_seg.jpg"  width="160"/> | <img src="./assets/collision_edges.jpg"  width="160"/> |
-| Box pickup   | <img src="./assets/scenario_box_pickup.jpg" width="160"/> | <img src="./assets/box_pickup_depth.jpg" width="160"/> | <img src="./assets/box_pickup_segmentation.jpg" width="160"/> | <img src="./assets/box_pickup_shaded_seg.jpg" width="160"/> | <img src="./assets/box_pickup_edges.jpg" width="160"/> |
+The synthetic origin of the dataset gives us access to deterministic, perfectly-aligned ground truth for every frame, rendered directly by the simulator rather than predicted by a model. The figure below shows, for a single representative frame from each scenario, the RGB video together with the four annotation modalities that are visible as imagery: log-normalized colorized metric depth, instance segmentation (colorized so the per-pixel identity is visible), shaded segmentation (the same per-pixel identity rendered with normal-based shading), and a Canny edge map computed on the shaded segmentation.
 
----
+![Ground-truth modalities — one frame per scenario across RGB, depth, instance segmentation, shaded segmentation, and Canny edges.](./assets/modalities_grid.png)
 
-## Dataset statistics
+In addition to the imagery shown above, every frame ships with per-agent two-dimensional axis-aligned bounding boxes (both tight and loose), per-agent oriented three-dimensional bounding boxes, and the camera intrinsics and extrinsics that produced the frame. These structured annotations live in per-camera consolidated JSON files in the upcoming artifacts tier.
 
-| Scenario                     | # clips    | # runs (samples) | Clip length | Cameras / run | Repo path |
-|------------------------------|-----------:|-----------------:|------------:|--------------:|---|
-| 🚜 Forklift–human near-miss  | 27,939     | 13,410 *(¹)*     | 10 s        | 10 / 1        | `rgb/forklift_human_nearmiss/` |
-| 🔥 Warehouse fire            | 44,734     | 9,064            | 10 s        | 5             | `rgb/warehouse_fire/` |
-| 💥 Forklift–shelf collision  | 24,617     | 4,120            | 15 s        | 6             | `rgb/forklift_shelf_collision/` |
-| 📦 Warehouse box pickup      | 25,677     | 2,601            | 15 s        | 10            | `rgb/warehouse_box_pickup/` |
-| **Total**                    | **122,967**| **29,195**       | —           | —             | — |
+The current RGB tier publishes the photoreal RGB video and the run-level metadata. The artifacts tier publishes the four modality videos shown above (`depth.mp4`, `segmentation.mp4`, `shaded_seg.mp4`, and `edges.mp4`), the per-frame raw instance-segmentation PNGs that contain the underlying integer identities, the per-camera consolidated `camera_params.json` and `object_detection.json` files, and the run-level IRO randomization configuration. The shard keys are designed so that the RGB tier and the artifacts tier join cleanly on the per-sample `run_id`.
 
-Aggregate footage: **≈ 412 hours** at 1920 × 1080 / 30 fps. The "samples" column is the number of `__key__` entries you'll see when iterating with WebDataset (i.e. tar groups, not individual mp4 entries).
+## Dataset statistics
 
-> *(¹)* **Nearmiss composition.** Nearmiss is a **union of two source pipelines**:
-> - **1,642 multi-camera runs** from the IRA/IRO multi-cam pipeline (10 cameras each → 16,171 clips, with `source_kind = "multi_camera_run"` in `meta.json`).
-> - **11,768 single-camera clips** legacy-merged from earlier `warehouse_5k_videos/` and `warehouse_6k_videos/` training-set dumps (one camera each, packed as 1-camera samples with `source_kind = "extras_single_view"`, no per-run scene composition metadata). These are listed at the END of the nearmiss shard range (`nearmiss-rgb-00051.tar` onward).
->
-> Filter by `meta["source_kind"]` or by `clips.parquet["source_kind"]` to select one or both.
+| Scenario | Number of clips | Number of runs (WebDataset samples) | Clip length | Cameras per run | Repository path |
+|---|---:|---:|---:|---:|---|
+| Forklift–human near-miss | 27,939 | 13,410 | 10 seconds | 10 or 1 | `rgb/forklift_human_nearmiss/` |
+| Warehouse fire | 44,734 | 9,064 | 10 seconds | 5 | `rgb/warehouse_fire/` |
+| Forklift–shelf collision | 24,617 | 4,120 | 15 seconds | 6 | `rgb/forklift_shelf_collision/` |
+| Warehouse box pickup | 25,677 | 2,601 | 15 seconds | 10 | `rgb/warehouse_box_pickup/` |
+| **Total** | **122,967** | **29,195** | — | — | — |
 
----
+The aggregate footage is approximately 412 hours at 1920 by 1080 resolution and 30 frames per second. The "Number of runs" column corresponds to distinct WebDataset samples, that is, the number of `__key__` values you will observe when iterating with a WebDataset reader.
 
-## Simulation infrastructure
+The near-miss row is heterogeneous because it merges two source pipelines. The first is 1,642 multi-camera runs from the Isaac Replicator Object plus Isaac Replicator Agent pipeline, each captured from ten synchronized cameras, contributing 16,171 clips. The second is 11,768 single-camera clips that were generated by an earlier pipeline and that we include here for completeness, packed as one-camera samples with no per-run scene-composition metadata. The two sources are distinguished by the `source_kind` field in each sample's `meta.json` and in `metadata/clips.parquet`, with the values `multi_camera_run` and `extras_single_view` respectively. The single-camera near-miss samples are located in the shard range `nearmiss-rgb-00051.tar` through `nearmiss-rgb-00112.tar`.
 
-All four scenarios are built on **NVIDIA Isaac Sim**:
+## Simulation pipeline
 
-- **Procedural scene composition** — warehouse layout, shelf placement, prop variation, and per-light randomization of color temperature, intensity, exposure, and color — handled by **Isaac Sim Replicator Object (IRO)**.
-- **Agent and sensor population** — worker spawning and behavior, forklift placement and navigation, and the camera rigs that define the dataset's multi-view viewpoints — handled by **Isaac Sim Replicator Agent (IRA)**.
-- **Camera placement** is parametric, with height, distance, and look-down angle sampled per run.
-- **Worker assets and motions** are sampled from Isaac Sim's character library to diversify human appearance and gait.
-- Each simulation run is **seeded with a unique random seed** (exposed as `meta.seed` per sample) that controls all randomized variables (scene composition, lighting, agent identity and motion, camera pose, event timing), so runs are independent and reproducible.
+All four scenarios are built on NVIDIA Isaac Sim. Procedural scene composition — warehouse layout, shelf placement, prop variation, and per-light randomization of color temperature, intensity, exposure, and color — is handled by the Isaac Sim Replicator Object component. Agent and sensor population — worker spawning and behavior, forklift placement and navigation, and the camera rigs that define the dataset's multi-view viewpoints — is handled by the Isaac Sim Replicator Agent component. Camera placement is parametric, with height, distance, and look-down angle sampled per run. Worker assets and motions are sampled from Isaac Sim's character library to diversify human appearance and gait.
 
----
+Each simulation run is seeded with a unique random seed that controls every randomized variable: scene composition, lighting, agent identity, agent motion, camera pose, and event timing. The seed is exposed in each sample's `meta.json` (and in the Parquet indexes), so any individual run is fully reproducible from this dataset alone, and the same pipeline can be extended to additional scenarios outside this release without modification.
 
 ## Repository layout
 
 ```
 nvidia/PhysicalAI-SDG-WareHouse/
 ├── README.md
+├── assets/                     ← images used by this dataset card
 ├── metadata/
-│   ├── runs.parquet           # 1 row per run/sample, with clip_keys[], shard_path, etc.
-│   ├── clips.parquet          # 1 row per (run × camera), with scenario, seed, hash_filename, …
-│   └── manifests/             # provenance copies of the source-S3 manifests
-├── rgb/                       # ← this release
-│   ├── forklift_human_nearmiss/
-│   │   ├── nearmiss-rgb-00000.tar
-│   │   ├── nearmiss-rgb-00001.tar
-│   │   └── … (113 shards, ~5 GiB each)
-│   ├── warehouse_fire/        # 125 shards
-│   ├── forklift_shelf_collision/   # 114 shards
-│   └── warehouse_box_pickup/  # 107 shards
-└── artifacts/                 # ← coming soon: depth/seg/edges + bboxes + camera params
+│   ├── runs.parquet            ← one row per WebDataset sample (run), with
+│   │                             scenario, seed, source_kind, n_cameras,
+│   │                             total_bytes, shard_path_in_repo, clip_keys
+│   ├── clips.parquet           ← one row per (run × camera), with hash_filename,
+│   │                             camera_alias, source_rgb_s3, size, etc.
+│   └── manifests/              ← provenance copies of the source-S3 manifests
+└── rgb/                        ← this release
+    ├── forklift_human_nearmiss/
+    │   ├── nearmiss-rgb-00000.tar
+    │   ├── nearmiss-rgb-00001.tar
+    │   └── …                   (113 shards, ~5 GiB each)
+    ├── warehouse_fire/         (125 shards)
+    ├── forklift_shelf_collision/   (114 shards)
+    └── warehouse_box_pickup/   (107 shards)
 ```
 
-Each `.tar` is a **WebDataset** archive. Inside it, every sample (= one simulation run = one set of synchronized cameras) is a group of entries sharing the same stem (`run_id`):
+Each `.tar` is a WebDataset archive. Inside it, every sample is a group of entries that share the same stem, where the stem is the `run_id`:
 
 ```
 fd7cc35596b247b16b0b_run_8_seed_864110064.meta.json
@@ -272,16 +165,17 @@ fd7cc35596b247b16b0b_run_8_seed_864110064.eye_00.rgb.mp4
 …
 ```
 
-WebDataset readers will yield one Python dict per `run_id` with `__key__ = run_id` and keys `meta.json`, `metadata.txt` (when present), and one `{alias}.rgb.mp4` per camera.
-
----
+WebDataset readers yield one Python dictionary per run, with `__key__` set to the `run_id`, a `meta.json` entry, an optional `metadata.txt` entry, and one `{camera_alias}.rgb.mp4` entry per camera.
 
 ## Loading examples
 
-### Stream from HF without downloading shards — WebDataset
+### Stream directly from the Hub with WebDataset
+
+The following example streams the fire scenario directly from the Hub using the standard WebDataset reader. Nothing is materialized on disk apart from the bytes that the iterator actually consumes.
 
 ```python
-import os, webdataset as wds
+import os
+import webdataset as wds
 from huggingface_hub import get_token
 
 token = get_token() or os.environ["HF_TOKEN"]
@@ -292,17 +186,19 @@ url = (
     f"-H 'Authorization: Bearer {token}'"
 )
 
-ds = wds.WebDataset(url, shardshuffle=True).decode()
-for sample in ds.shuffle(1000):
+dataset = wds.WebDataset(url, shardshuffle=True).decode()
+for sample in dataset.shuffle(1000):
     run_id = sample["__key__"]
-    meta   = sample["meta.json"]          # dict
-    # one bytes blob per camera:
-    rgb    = sample["ceiling_00.rgb.mp4"] # bytes, decode with av/ffmpeg
-    print(run_id, meta["seed"], list(k for k in sample if k.endswith(".rgb.mp4")))
+    meta = sample["meta.json"]            # dict: scenario, seed, cameras, etc.
+    rgb_bytes = sample["ceiling_00.rgb.mp4"]  # raw mp4 bytes; decode with av/ffmpeg
+    camera_keys = sorted(k for k in sample if k.endswith(".rgb.mp4"))
+    print(run_id, meta["seed"], camera_keys)
     break
 ```
 
-### Filter with Parquet, then download just the shards you need
+### Filter with the Parquet index, then fetch only the shards you need
+
+The Parquet indexes let you select clips or runs by any combination of scenario, camera alias, source kind, or seed, and recover the exact shard paths to fetch.
 
 ```python
 import pandas as pd
@@ -316,13 +212,14 @@ clips = pd.read_parquet(
     )
 )
 
-# All eye-level views from fire runs with even seed:
-sel = clips[
+# All eye-level views from fire runs with an even seed:
+selection = clips[
     (clips.scenario == "fire")
     & (clips.camera_alias.str.startswith("eye_"))
     & (clips.seed.notna()) & (clips.seed % 2 == 0)
 ]
-print(sel.shape, sel.shard_path_in_repo.nunique(), "unique shards to fetch")
+unique_shards = sorted(selection.shard_path_in_repo.unique())
+print(f"{len(selection):,} clips across {len(unique_shards)} shards")
 ```
 
 ### Pull one scenario only with the CLI
@@ -338,6 +235,7 @@ huggingface-cli download nvidia/PhysicalAI-SDG-WareHouse \
 
 ```python
 from huggingface_hub import hf_hub_download
+
 local_tar = hf_hub_download(
     repo_id="nvidia/PhysicalAI-SDG-WareHouse",
     repo_type="dataset",
@@ -345,44 +243,37 @@ local_tar = hf_hub_download(
 )
 ```
 
----
-
 ## Dataset description
 
 | Field | Value |
 |---|---|
-| **Owner** | NVIDIA |
-| **Creation date** | 2026 |
-| **Data collection method** | Synthetic (NVIDIA Isaac Sim, IRO + IRA) |
-| **Labeling method** | Automatic (Isaac Sim Replicator) |
-| **Container / codec** | MP4 (H.264) |
-| **Resolution** | 1920 × 1080 |
-| **Frame rate** | 30 FPS |
-| **Packaging** | WebDataset (tar shards, ~5 GiB each) |
-| **Language(s) of metadata** | English |
-| **License** | CC BY 4.0 |
+| Owner | NVIDIA |
+| Creation date | 2026 |
+| Data collection method | Synthetic (NVIDIA Isaac Sim, with the Isaac Replicator Object and Isaac Replicator Agent components) |
+| Labeling method | Automatic (Isaac Sim Replicator) |
+| Container and codec | MP4 (H.264) |
+| Resolution | 1920 × 1080 |
+| Frame rate | 30 frames per second |
+| Packaging | WebDataset tar shards, approximately 5 GiB each |
+| Metadata language | English |
+| License | CC BY 4.0 |
 
----
+## Known limitations and future work
 
-## Known limitations
+This release publishes the RGB tier only. The full annotation stack — metric depth, instance and shaded segmentation, Canny edges, two-dimensional and three-dimensional bounding boxes, and per-frame camera intrinsics and extrinsics — is already rendered, aligned to every RGB frame, and will be added to this same repository as a companion `artifacts/` tier. The visualizations in the ground-truth modalities figure above were produced from those native renders.
 
-- **v1 = RGB only.** The annotation stack (depth, instance + shaded segmentation, Canny edges, 2D/3D bounding boxes, per-frame camera intrinsics+extrinsics) is rendered and aligned to every RGB frame but ships in an upcoming `artifacts/` tier of this same repo. The visualization gallery above shows those native renders for reference.
-- **Heterogeneous nearmiss composition.** Nearmiss combines 1,642 multi-camera runs (10 cameras each) with 11,768 single-camera legacy clips from older training-set dumps; the latter lack run-level scene-composition metadata and live in shards `nearmiss-rgb-00051..00112`. Use `meta["source_kind"]` to disambiguate.
-- **Sim-to-real gap.** SDG-Warehouse is fully synthetic and may exhibit appearance differences from real warehouse footage, including a CG-like look, simplified material response, and limited fidelity in volumetric effects such as smoke and fire.
-- **Stylized agent motion.** Worker motion is driven by procedural navigation and behavior, so reactions — particularly emergency evacuation in the fire scenario — can occasionally appear unnatural.
-- **Approximate fine-grained physics.** Rigid-body interactions (forklift contact with shelves and props) are physically simulated, but very fine-grained debris, deformation, and secondary contact effects are approximate.
-- **Environment diversity.** The release is concentrated on a single warehouse layout family; future work will add additional warehouse, retail, and factory floor plans.
-- **Incident coverage.** Future work will add event types such as spills, dropped pallets, and shelf collapses without forklift involvement.
-- **Human-attribute diversity.** Future work will broaden variation in worker attire and personal protective equipment (PPE).
+The near-miss scenario is heterogeneous in composition. It combines 1,642 multi-camera runs (ten cameras each) with 11,768 single-camera legacy clips from older training-set dumps; the latter do not carry run-level scene-composition metadata and are located in the second half of the near-miss shard range. The `source_kind` field on every sample disambiguates the two sources.
 
----
+The dataset is fully synthetic and exhibits a sim-to-real gap. Compared to real warehouse footage, the rendered material can have a computer-graphics-like appearance, simplified material response, and limited fidelity in volumetric effects such as smoke and fire. Models trained on the dataset should be carefully evaluated on real footage before being deployed in any safety-critical setting.
+
+Agent motion is driven by procedural navigation and behavior, so reactions — and in particular the coordinated evacuation behavior in the fire scenario — can occasionally appear unnatural. Rigid-body interactions between the forklift and the storage shelves are physically simulated, but very fine-grained debris, deformation, and secondary contact effects are approximate.
+
+The current release is concentrated on a single warehouse layout family. Future work will broaden environment diversity to additional warehouse, retail, and factory floor plans, will add additional incident types such as spills, dropped pallets, and shelf collapses without forklift involvement, and will broaden variation in worker attire and personal protective equipment.
 
 ## Contributors
 
 Nalin Dadhich, Jiajun Li, Robert Denomme, Prahan Reddy Poreddy, Patrick Kim.
 
----
-
 ## Citation
 
 If you use SDG-Warehouse in your research, please cite the Cosmos3 technical report:
@@ -397,18 +288,14 @@ If you use SDG-Warehouse in your research, please cite the Cosmos3 technical rep
 }
 ```
 
----
-
 ## License
 
-Released under the **Creative Commons Attribution 4.0 International (CC BY 4.0)** license.
-
----
+Released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
 
 ## Ethical considerations
 
-NVIDIA believes Trustworthy AI is a shared responsibility, and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
+NVIDIA believes trustworthy AI is a shared responsibility, and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this dataset meets the requirements for their relevant industry and use case, and addresses any unforeseen product misuse.
 
-SDG-Warehouse is **fully synthetic** — no real people, real workplaces, or real surveillance footage are included — and depicts safety-critical events (near-miss, collision, fire, evacuation) only in simulation. Models trained on it should still be carefully evaluated on real footage before being deployed in any safety-critical setting, and operators should be aware of the sim-to-real gap noted above.
+SDG-Warehouse is fully synthetic. It contains no real people, no real workplaces, and no real surveillance footage, and it depicts safety-critical events — near-misses, collisions, fires, and evacuations — only in simulation. Models trained on it should still be carefully evaluated on real footage before being deployed in any safety-critical setting, and operators should be aware of the sim-to-real gap noted above.
 
 Please report security vulnerabilities or NVIDIA AI concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).