Card: SuperOdom-only GT, real-robot data, add structure/sample figures

README.md

@@ -29,15 +29,18 @@ humanoid benchmark in *Self-Supervised Inertial Odometry* (SSIO, Qiu et al., 202
 - 14 body parts per clip (pelvis, head, arms, hands, legs, feet), each with an IMU stream, a world
   pose and a body-frame velocity.
 
+## Overview
+
+![Data structure: per-clip files, channel layout, and the 14 body-part indices](assets/data_structure.png)
+
+![Test clip: pelvis ground-truth trajectory, IMU and ground-truth velocity](assets/sample_clip.png)
+
 ## Collection platform and ground truth
 
 - Robot: Unitree G1 humanoid.
-- Ground truth: LiDAR-inertial odometry (SuperOdom). The SSIO training code refers to these clips as
-  "SuperOdometry-GT clips". The paper also mentions learning-based stereo VO (MAC-VO); see TODO
-  below on which source(s) these particular files contain.
-- Each `spec.json` has `"kind": "sim"`, and the SSIO loader says the clips "store synthesized
-  body-part IMUs". The 14 per-part IMU streams are therefore **synthesized**. See TODO on how they
-  are made and whether the pelvis stream is the physical G1 IMU.
+- Ground truth: LiDAR-inertial odometry (SuperOdom).
+- All clips are real-robot recordings. The `"kind"` field in `spec.json` is a converter tag, not a
+  marker of simulated data.
 
 ## Directory layout
 
@@ -47,6 +50,7 @@ humanoid benchmark in *Self-Supervised Inertial Odometry* (SSIO, Qiu et al., 202
     measure.npy               # (P, T, 9) float16  [acc | gyro | gt_vel]
     pose.npy                  # (P, T, 7) float32  world pose (tx, ty, tz, qx, qy, qz, qw)
     time_ns.npy               # (T,)      int64    timestamps in ns, shared by all parts
+assets/                       # card figures
 splits/
     train.txt                 # 31 clip names
     test.txt                  # 1 clip name
@@ -193,12 +197,7 @@ The SSIO reference loader is `SSIO/Sequence/HumanoidOdom.py`. It uses the pelvis
 ## TODO before release
 
 - **License**: `license_name: tbd`. Pick a license.
-- **Ground-truth source per clip**: the code says SuperOdom (LiDAR-inertial). Check whether any
-  clip uses MAC-VO stereo VO instead of, or fused with, SuperOdom, and give the GT rate before
-  interpolation to the IMU grid.
-- **Synthesized IMUs** (`"kind": "sim"`): describe how the 14 body-part IMU streams are generated,
-  for example from GT kinematics plus noise, or transferred from the physical IMU. State whether the
-  `pelvis` stream is the real G1 IMU. Give the noise model and the IMU model.
+- **Ground-truth rate**: give the SuperOdom output rate before interpolation to the IMU grid.
 - **World frame `W`**: z-up is verified, but the origin and heading are not. The pelvis starts at
   roughly `z ≈ -0.5 m` with near-identity orientation, which suggests the origin sits at a sensor
   above the pelvis. Confirm that.