Upload README.md with huggingface_hub

README.md

@@ -2,7 +2,7 @@
 license: cc-by-4.0
 language:
 - en
-pretty_name: Ego-Tactile Manipulation (Action-Segmented)
+pretty_name: Ego-Tactile Manipulation
 size_categories:
 - 100K<n<1M
 task_categories:
@@ -15,127 +15,79 @@ tags:
 - tactile
 - manipulation
 - action-segmentation
-- imitation-learning
 - multimodal
 ---
 
-# Ego-Tactile Manipulation — Action-Segmented
+# Ego-Tactile Manipulation
 
-**First-person video + dense worn-glove tactile + tactile-grounded action segmentation** for bimanual tabletop manipulation. Every frame carries the egocentric camera, **160 channels of fingertip pressure** (both hands), and a physically-grounded action label — released in [**LeRobot v3.0**](https://huggingface.co/docs/lerobot/en/lerobot-dataset-v3) format with ready-to-open [**Rerun**](https://rerun.io) visualizations.
+*First-person video + dense two-hand tactile + touch-grounded action labels — by [**OpenGraph Labs**](https://www.opengraphlabs.com/).*
 
-Synchronized **ego + touch** human-manipulation data is scarce. Most egocentric datasets are vision-only; most tactile datasets are single-finger probes on a bench. This release pairs a head camera with a **full five-finger tactile glove on each hand**, and the action boundaries are derived from the **physical contact and grip-force signal** — not from human annotation — so the labels are grounded in what the hands actually did.
+![Ego-tactile capture and visualization](https://huggingface.co/datasets/OpenGraphLabs-Research/ego-tactile-manipulation/resolve/main/assets/ego_tactile.png)
 
-> A compact, clean **1.28-hour sample** from OpenGraph Labs' Physical-AI data pipeline. Built to be loaded and visualized in one line.
+Synchronized **ego + touch** human-manipulation data is rare. This is a clean **1.28-hour sample** from OpenGraph Labs' Physical-AI data pipeline: a head camera plus our **OGLO tactile gloves** on both hands, with action labels derived from the physical contact signal — not hand-annotated.
 
-## At a glance
+## OGLO — our tactile gloves
 
-| | |
-|---|---|
-| Episodes | **72** |
-| Duration | **1.28 hours** (138,629 frames @ 30 fps) |
-| Embodiment | Human, **bimanual**, egocentric (head camera + tactile gloves) |
-| Camera | iPhone ultra-wide, **1920×1080** @ 30 fps (intrinsics included) |
-| Tactile | **2 hands × 5 fingers × 4×4 taxels = 160 channels**, 30 fps aligned + **90 Hz raw** sidecar |
-| Annotations | Tactile-grounded **L0 action / L1 sub-action / anchor / grasp / hand / object** (English) |
-| Tasks | Diverse tabletop manipulation — pouring/dispensing condiments, opening & closing bottle caps, handling cups, boxes, erasers, packaging |
-| Format | **LeRobot v3.0** (`lerobot >= 0.4`) + **Rerun** `.rrd` |
-| License | **CC-BY-4.0** |
+![OGLO tactile gloves](https://huggingface.co/datasets/OpenGraphLabs-Research/ego-tactile-manipulation/resolve/main/assets/oglo_gloves.png)
 
-## Modalities (per frame)
+Worn on both hands · **5 fingers × 4×4 taxels = 80 channels per hand (160 total)** · 90 Hz pressure, co-registered to every video frame.
 
-| Feature | Type | Shape | Description |
-|---|---|---|---|
-| `observation.images.ego` | video | 1080×1920×3 | Egocentric RGB |
-| `observation.tactile.left` / `.right` | float32 | (80,) | Fingertip pressure, 5 fingers × 4×4 taxels, video-aligned 30 Hz |
-| `observation.wrist_imu.left` / `.right` | float32 | (6,) | Wrist accel (xyz) + gyro (xyz) |
-| `observation.imu.head` | float32 | (9,) | Head accel + gyro + gravity |
-| `annotation.l0_action` | string | — | The action (also the LeRobot `task`) |
-| `annotation.l1_subaction` | string | — | The tactile-grounded sub-step |
-| `annotation.anchor` | string | — | `contact_on` · `load` · `release` · `hand_switch` · `continuation` · `transition` |
-| `annotation.grasp_type` | string | — | `pinch` · `power` · `lateral` · `partial` |
-| `annotation.hand` | string | — | `left` · `right` · `both` |
-| `annotation.target_object` | string | — | Manipulated object |
+## Touch-grounded action segmentation
 
-## Tactile sensing
+Action boundaries come from physical **contact and grip-force** — no human ground truth. Concise labels are added within each physically-fixed span.
 
-Each hand wears a glove with a **5-finger × 4×4 taxel grid (80 channels/hand)**. The 80-vector block order is `index, middle, pinky, ring, thumb`; within a finger, channel `i → row = (i%16)//4, col = (i%16)%4` (see `names` in `meta/info.json`).
+![Action storyboard](https://huggingface.co/datasets/OpenGraphLabs-Research/ego-tactile-manipulation/resolve/main/assets/action_seg_storyboard.png)
+![Action browser](https://huggingface.co/datasets/OpenGraphLabs-Research/ego-tactile-manipulation/resolve/main/assets/action_seg_browser.png)
 
-- **Aligned (30 Hz):** `observation.tactile.{left,right}` — co-registered to every video frame, ready for training.
-- **Raw (90 Hz, byte-exact):** `extra/raw_tactile/episode_*_{left,right}.parquet` — full-rate capture with `capture_ns` timestamps and the same channel names, for sub-frame contact dynamics.
+## Specs
 
-## Action segmentation — grounded in touch, no human labels
+| | |
+|---|---|
+| Episodes / duration | 72 · **1.28 h** (138,629 frames @ 30 fps) |
+| Camera | egocentric, 1920×1080 (intrinsics included) |
+| Tactile | 160 channels · 30 Hz aligned + 90 Hz raw |
+| IMU | head (9-ch) + both wrists (6-ch each) |
+| Annotations | action / sub-action / anchor / grasp / hand / object (English) |
+| Tasks | tabletop manipulation — pour, cap, cut, grip, arrange, … |
+| License | CC-BY-4.0 |
 
-Action boundaries come from the **physical tactile signal**, not from human annotation:
+## What's inside
 
-- **L0 (action):** a contact episode — any-hand contact onset → offset.
-- **L1 (sub-action):** cut within an L0 by **hand-role changes** (L/R/both) and **salient grip-force events** (load / release), each tagged with the physical **anchor** that produced it.
+```
+meta/      LeRobot v3.0 metadata  (+ opengraph/: camera intrinsics, tactile highlights)
+data/      per-frame parquet      (tactile @30 Hz, IMU, annotations)
+videos/    egocentric video
+extra/     raw 90 Hz tactile      (byte-exact)
+rerun/     one .rrd per episode   (instant visualization)
+```
 
-Within these physically-fixed spans a vision-language model assigns concise labels constrained by the contact context — so the language describes what the hand is verifiably doing. Per-episode segmentation metadata (anchors, grip-force, key-frames) is preserved in `meta/opengraph/tactile_highlights/`.
+## Format & use
 
-## Load it (LeRobot)
+Packaged in **[LeRobot v3.0](https://huggingface.co/docs/lerobot)** and **[Rerun](https://rerun.io)**, so it's ready to use as-is:
 
 ```python
-# pip install "lerobot>=0.4"
 from lerobot.datasets import LeRobotDataset
-
 ds = LeRobotDataset("OpenGraphLabs-Research/ego-tactile-manipulation")
-print(ds.num_episodes, ds.num_frames)          # 72  138629
-
-sample = ds[0]
-sample["observation.images.ego"]               # (3, 1080, 1920)
-sample["observation.tactile.left"]             # (80,)
-sample["annotation.l1_subaction"]              # e.g. "tilt the ketchup bottle"
-sample["task"]                                 # the L0 action
 ```
-
-## Visualize it (Rerun)
-
-Two ways:
-
 ```bash
-# 1) Tailored, self-contained recordings shipped with the dataset
-#    (ego camera | left/right per-finger grip charts | action-segmentation log)
-pip install rerun-sdk
-rerun rerun/episode_000000.rrd
-
-# 2) Stream any episode straight from the LeRobot dataset
-lerobot-dataset-viz --repo-id OpenGraphLabs-Research/ego-tactile-manipulation --episode-index 0
+rerun rerun/episode_000000.rrd      # visualize an episode
 ```
 
-## Directory layout
-
-```
-meta/                       LeRobot v3.0 metadata (info.json, stats, tasks, episodes)
-  opengraph/                OpenGraph extensions:
-    camera_intrinsics/      per-episode camera K (fx, fy, cx, cy, distortion)
-    tactile_highlights/     per-episode L0/L1 anchors + grip-force + key-frames
-data/                       frame-level parquet (tactile @30Hz, IMU, annotations)
-videos/                     egocentric video (LeRobot-encoded)
-extra/raw_tactile/          raw 90 Hz tactile, byte-exact (capture_ns + 80 ch)
-rerun/                      one .rrd per episode — instant visualization
-```
+## About OpenGraph Labs
 
-## Limitations & scope
+Physical-AI data infrastructure — we capture and post-process large-scale egocentric human manipulation data (multi-camera, tactile, IMU, metric egomotion, frame-level semantics). This is our first public sample; reach out for collaborations and larger releases.
 
-- **Sample release (1.28 h).** A clean slice of a larger pipeline — meant to demonstrate the ego+tactile+annotation stack end-to-end.
-- **No 3D head/hand pose** in this release (the egomotion/hand-pose stage was not run for these episodes).
-- **Auto-annotated.** Boundaries are physically grounded; the language labels are model-generated and constrained by contact context — not exhaustively human-verified.
-- Tabletop manipulation in a controlled capture setting; gloved hands.
+🌐 [opengraphlabs.com](https://www.opengraphlabs.com/) · 🤗 [OpenGraphLabs-Research](https://huggingface.co/OpenGraphLabs-Research)
 
-## License & citation
-
-Released under **CC-BY-4.0** — free to use, including commercially, with attribution.
+## Citation
 
 ```bibtex
 @misc{opengraphlabs2026egotactile,
-  title  = {Ego-Tactile Manipulation: Egocentric Video with Dense Tactile and
-            Tactile-Grounded Action Segmentation},
+  title  = {Ego-Tactile Manipulation},
   author = {OpenGraph Labs},
   year   = {2026},
   url    = {https://huggingface.co/datasets/OpenGraphLabs-Research/ego-tactile-manipulation}
 }
 ```
 
-## About
-
-Built by **[OpenGraph Labs](https://huggingface.co/OpenGraphLabs-Research)** — Physical-AI data infrastructure. We capture and post-process large-scale egocentric human manipulation data (multi-camera, tactile, IMU, metric egomotion, frame-level semantics). This is our first public sample; reach out for collaborations and larger releases.
+Released under **CC-BY-4.0** — free to use, including commercially, with attribution.