Add outdoor split, loader-schema docs, and refresh the dataset card

DATA_README.md

@@ -0,0 +1,137 @@
+# MIPO
+
+Data was recorded on a Unitree Go1 with 4 auxilliary IMUs mounted near the feet on the calf links. It also has joint states and foot contact sensors. Indoor trials have OptiTrack motion-capture ground truth. Outdoor trials have no motion capture. The original data is from the [Cerberus2.0 repo](https://github.com/ShuoYangRobotics/Cerberus2.0) and the "Multi-IMU Proprioceptive Odometry for Legged Robots" related paper, hence the adopted shorthand, MIPO. That data was saved as a rosbag (indoor) or rosbag + phone recording (outdoor) and can be retrieved on that repo through the linked google drive. For ease of use, we converted this data to hdf5 format. In the HDF5 files, each per-topic stream is stored at its native rate with its own ROS timestamp axis. Synchronization, resampling, filtering, and calibration are not applied.
+
+**On loading.** The files are intentionally unsynchronized and uncalibrated: no per-topic
+resampling, no kalibr time-offset or extrinsic applied, no phone-to-robot clock correction.
+Reading the groups directly with `h5py` will silently skip all of that and produce a result that
+looks plausible but is misaligned. It is recommended to use, or faithfully reapply, the
+calibration from the [SSIO loader](https://github.com/SSIO-IMU/SSIO)
+(`Sequence.MIPO.MIPO_Sequence`, the `MIPO_*` dataset configs), which already implements it
+correctly.
+
+## Splits
+
+- **`indoor/`**: lab recordings with OptiTrack ground truth. See [`indoor/README.md`](indoor/README.md) for schema and calibration.
+- **`outdoor/`**: field recordings with no robot ground truth. A position reference is reconstructed from an operator-carried phone's GPS + IMU. See [`outdoor/README.md`](outdoor/README.md) for schema, calibration, and what this means for "ground truth."
+
+Each split is calibrated by its own method: indoor via kalibr-estimated time offset + T_BS
+extrinsic against OptiTrack, and outdoor via a phone-to-robot clock offset and a fixed onboard-AHRS
+tilt correction.
+
+## Robot sensor data
+
+Both splits record the robot side with the same sensors, schema, and rates (confirmed directly
+against the files, not just assumed), since it's the same Go1 setup in both cases. Each split's
+README covers what's different: ground truth, calibration, and the split-specific exceptions noted
+below each table.
+
+Group paths mirror the ROS topic tree (but with leading `/` stripped). Every leaf is a contiguous
+numeric array (time-major along axis 0). Covariances are flattened row-major and are also the
+zero-placeholders that were present in the rosbags to comply with `sensor_msgs/Imu` format. I.e.
+the HDF5 is meant to mirror ROS formatting without rosbag storage and access overhead. All `time`
+arrays are ROS message stamps in POSIX seconds (float64).
+
+### Root attributes (each `.h5`)
+
+| Attribute | Type | Meaning |
+|---|---|---|
+| `imu_topics` | JSON-encoded string list | The 5 IMU topic names included. |
+| `joint_topic` | string | `/unitree_hardware/joint_foot` |
+| `num_imus` | int | 5 (1 main-body + 4 aux near each foot). |
+| `message_counts` | JSON-encoded dict | `{topic: count}` for every preserved topic. |
+| `source_bag_rel` | string | Source bag path relative to the dataset root. Bags not in this release. |
+| `compression` | string | `gzip`. |
+| `created_utc` | string | ISO-8601 export timestamp. |
+
+Indoor additionally has a `gt_topic` attribute (`/natnet_ros/Shuo_Go1/pose`). Outdoor has none, see
+`outdoor/README.md`.
+
+### `unitree_hardware/imu/` (body IMU, ~400 Hz)
+
+`sensor_msgs/Imu` schema.
+
+| Dataset | Shape | Units |
+|---|---|---|
+| `linear_acceleration` | (N, 3) | m/s², sensor body frame, gravity included |
+| `angular_velocity` | (N, 3) | rad/s, sensor body frame |
+| `orientation` | (N, 4) | quaternion (x, y, z, w), Unitree onboard AHRS. Population varies by sequence, see each split's README. |
+| `linear_acceleration_cov` | (N, 9) | row-major 3×3 (all zero) |
+| `angular_velocity_cov` | (N, 9) | row-major 3×3 (all zero) |
+| `orientation_cov` | (N, 9) | row-major 3×3 (all zero) |
+| `time` | (N,) | sec |
+
+### `WT901_47_Data/`, `WT901_48_Data/`, `WT901_49_Data/`, `WT901_50_Data/` (calf-mounted IMUs, ~333 Hz)
+
+Four WitMotion WT901 IMUs mounted on the calfs near the robot feet. Same field set as the body
+IMU. `*_cov` fields are all zero. `orientation` is also all zero on every sequence (no AHRS).
+
+### `unitree_hardware/joint_foot/` (joints + foot contact, ~400 Hz)
+
+16 channels = 12 actuated joints + 4 foot-contact channels. These are taken directly in the format
+of the ROS bag, which are likely from the unitree SDK directly. The following are infered from
+that:
+- The foot contact channels are integer force count values (still stored as floats). This
+  motivates explicit normalization. This is applied in the MIPO Dataset loader using the stored
+  `mipo_joint_stats.json` precomputed stats from the train set. These are also applied for outdoor
+  data.
+- These force values are saved in the `effort` fields of the last 4 channels of the joints topic.
+
+| Dataset | Shape | Notes |
+|---|---|---|
+| `names` | (16,) | Joint names, in storage order. |
+| `position` | (N, 16) | rad (cols 12–15 zero) |
+| `velocity` | (N, 16) | rad/s (cols 12–15 zero on most bags, see each split's README for exceptions) |
+| `effort` | (N, 16) | cols 0–11 = joint torques (N·m), cols 12–15 = raw foot-force counts (integer-valued). See each split's README for the observed range. |
+| `time` | (N,) | sec |
+
+Joint name → meaning:
+
+| Idx | Name | Joint |
+|---|---|---|
+| 0 | FL0 | Front-left hip abduction/adduction |
+| 1 | FL1 | Front-left thigh (hip pitch) |
+| 2 | FL2 | Front-left calf (knee) |
+| 3–5 | FR0/1/2 | Front-right hip-abad / thigh / calf |
+| 6–8 | RL0/1/2 | Rear-left hip-abad / thigh / calf |
+| 9–11 | RR0/1/2 | Rear-right hip-abad / thigh / calf |
+| 12 | FL_foot | Front-left foot (force only, `effort[:, 12]`) |
+| 13 | FR_foot | Front-right foot (force only, `effort[:, 13]`) |
+| 14 | RL_foot | Rear-left foot (force only, `effort[:, 14]`) |
+| 15 | RR_foot | Rear-right foot (force only, `effort[:, 15]`) |
+
+**Outdoor ground truth is not true ground truth.** The outdoor position reference comes from a phone carried near the robot by the operator, not mounted on it. Use it as a coarse position and visual-comparison reference, not for velocity, not for high-accuracy position, and not as a substitute for true ground truth as network input. See `outdoor/README.md` for details.
+
+**Dropping bad data.** Sequences where calibration could not be trusted or another data-quality issue was flagged were excluded.
+
+## License and Attribution
+
+The recordings come from [Cerberus2.0](https://github.com/ShuoYangRobotics/Cerberus2.0)
+(AGPL-3.0) by Shuo Yang, Zixin Zhang, Benjamin Bokser, and Zachary Manchester. Please cite their
+work:
+
+```bibtex
+@inproceedings{yang2023mipo,
+      title={Multi-IMU Proprioceptive Odometry for Legged Robots},
+      author={Yang, Shuo and Zhang, Zixin and Bokser, Benjamin and Manchester, Zachary},
+      booktitle={IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
+      year={2023},
+}
+```
+
+This release is the quadruped benchmark used by [SSIO](https://ssio-imu.github.io)
+([paper](https://arxiv.org/abs/2610.25988), [code](https://github.com/SSIO-IMU/SSIO)), which
+trains inertial odometry from the robot's own unlabeled IMU stream. If you use this release, please
+also cite:
+
+```bibtex
+@misc{qiu2026ssioselfsupervisedinertialodometry,
+      title={SSIO: Self-Supervised Inertial Odometry},
+      author={Yuheng Qiu and Yutian Chen and Carlton Cort and Weison Huang and Shibo Zhao and Pieter Abbeel and Rocky Duan and Koushil Sreenath and C. Karen Liu and Carmelo Sferrazza and Junyi Geng and Sebastian Scherer and Guanya Shi},
+      year={2026},
+      eprint={2610.25988},
+      archivePrefix={arXiv},
+      primaryClass={cs.RO},
+      url={https://arxiv.org/abs/2610.25988},
+}
+```

README.md

@@ -20,15 +20,16 @@ size_categories:
 
 [Paper](https://arxiv.org/abs/2610.25988) | [Project Page](https://ssio-imu.github.io) | [Code](https://github.com/SSIO-IMU/SSIO)
 
-Indoor recordings of a Unitree Go1 quadruped: the body IMU, four foot-mounted IMUs, all
-12 joint states with 4 foot-force channels, and ground-truth pose from OptiTrack motion
-capture. It is the quadruped benchmark of **[SSIO](https://ssio-imu.github.io)**, which
-trains inertial odometry from the robot's own unlabeled IMU stream.
+Indoor and outdoor recordings of a Unitree Go1 quadruped: the body IMU, four foot-mounted IMUs,
+all 12 joint states with 4 foot-force channels, and a position reference for each split (OptiTrack
+motion capture GT indoors, a phone-based reference outdoors, see below). It is the quadruped
+benchmark of **[SSIO](https://ssio-imu.github.io)**, which trains inertial odometry from the
+robot's own unlabeled IMU stream.
 
-The recordings are from [Cerberus 2.0](https://github.com/ShuoYangRobotics/Cerberus2.0)
-and the paper *Multi-IMU Proprioceptive Odometry for Legged Robots* (Yang et al.,
-IROS 2023), hence the shorthand MIPO. This release converts the original ROS bags to
-HDF5, one file per bag, and adds the time-offset and extrinsic calibration that SSIO uses.
+The recordings are from [Cerberus 2.0](https://github.com/ShuoYangRobotics/Cerberus2.0) and the
+paper *Multi-IMU Proprioceptive Odometry for Legged Robots* (Yang et al., IROS 2023), hence the
+shorthand MIPO. This release converts the original ROS bags to HDF5, one file per bag, and adds
+the calibration SSIO uses.
 
 ## Dataset Structure
 
@@ -40,76 +41,57 @@ indoor/
 ├── 20230620_risqh_robustness_test/*.h5
 ├── 20230625_gait/*.h5
 └── 230622_robustness/*.h5
-splits/{train,valid,test}.txt
+outdoor/
+├── ahrs_gps_yaw_offsets.json
+├── <place>/*.h5                 (5 places: schedley park, mill19 trail,
+│                                 st mary catholic cemetery, frick park, CMU garage)
+└── pgo_densified_ref/<place>/*.npz
 ```
 
-Each `.h5` keeps its source bag's basename: `<date>-risqh-<gait>-<speed>-<height>-<path>`,
-with the gaits trot, flying trot (`flytrot`, `flyingtrot`), and stand trot.
-
-| Split | File | Sequences | Duration (min) |
-| --- | --- | ---: | ---: |
-| train | `splits/train.txt` | 36 | 28.9 |
-| valid | `splits/valid.txt` | 12 | 6.7 |
-| test | `splits/test.txt` | 12 | 10.0 |
-| **Total (all files)** | | **79** | **57.7** |
-
-The splits are the SSIO paper's and are disjoint. The 19 files outside the splits are
-kept for completeness; 5 of them have motion-capture freezes and jumps, so their ground
-truth is unusable (they are listed, commented out, in the SSIO `MIPO_Train` config).
-
-## Data Format
-
-Each topic is stored at its native rate with its own ROS timestamp axis (`time`, POSIX
-seconds, float64). No synchronization, resampling, filtering, or calibration is applied.
-Groups mirror the ROS topic tree:
-
-- **`unitree_hardware/imu/`** (~400 Hz): body IMU `linear_acceleration` (m/s², gravity
-  included), `angular_velocity` (rad/s), and onboard AHRS `orientation` (xyzw; all zero
-  on a few bags).
-- **`WT901_47_Data/` … `WT901_50_Data/`** (~333 Hz): the four WitMotion WT901 IMUs on the
-  calves, same fields as the body IMU (`orientation` is all zero).
-- **`unitree_hardware/joint_foot/`** (~400 Hz): `names`, `position`, `velocity`, `effort`
-  for 16 channels: 12 joints (`FL0..RR2`, rad, rad/s, N·m), then 4 feet whose `effort`
-  holds raw foot-force counts.
-- **`natnet_ros/Shuo_Go1/pose/`** (~60–70 Hz): OptiTrack `position` (m) and `orientation`
-  (xyzw, world ← body). No velocity stream.
-
-`indoor/mipo_time_offsets.json` holds per-sequence kalibr results (GT-to-IMU
-`time_offset_s`, `status`, `T_BS`) and a `_consensus` entry that the SSIO loader uses as
-the body-to-sensor extrinsic and fallback time offset. `indoor/mipo_joint_stats.json`
-holds the joint normalization statistics the SSIO checkpoints were trained with. The full
-schema is in
-[`Doc/mipo_hdf5_format.md`](https://github.com/SSIO-IMU/SSIO/blob/main/Doc/mipo_hdf5_format.md).
+`DATA_README.md` describes the robot sensor data shared by both splits (schema, units, rates) in
+full detail. Per-split specifics, ground truth, and calibration are in `indoor/README.md` and
+`outdoor/README.md`.
+
+**Indoor** has OptiTrack motion-capture ground truth. **Outdoor** has no motion capture; its
+position reference comes from a phone carried near the robot by the operator, not mounted on it.
+Use it as a coarse position and visual-comparison reference, not for velocity, not for
+high-accuracy position, and not as a substitute for true ground truth as network input.
+
+## Indoor Train/Valid/Test Split
+
+| Split | Sequences | Duration (min) |
+| --- | ---: | ---: |
+| train | 36 | 28.9 |
+| valid | 12 | 6.7 |
+| test | 12 | 10.0 |
+| **Total in pools** | **60** | **45.6** |
+| **Total indoor files (all)** | **79** | **57.7** |
+
+The splits are the SSIO paper's, disjoint, and reflect what the SSIO loader configs actually use.
+The 19 indoor files outside the splits are kept in this release for completeness; 5 have
+motion-capture freezes and jumps, so their ground truth is unusable, and 1 has frozen IMU data at
+the end of the sequence (see `indoor/README.md` for details on these 6).
 
 ## Quick Usage
 
 ```bash
-pip install huggingface_hub h5py
+pip install huggingface_hub
 hf download Open-IO/mipo-go1-odometry --repo-type dataset --local-dir mipo_hdf5
 ```
 
-```python
-from pathlib import Path
-import h5py
-
-seq = Path("mipo_hdf5/splits/train.txt").read_text().split()[0]
-with h5py.File(Path("mipo_hdf5/indoor") / seq) as f:
-    imu   = f["unitree_hardware/imu"]
-    acc, gyro, t_imu = imu["linear_acceleration"][:], imu["angular_velocity"][:], imu["time"][:]
-    gt    = f["natnet_ros/Shuo_Go1/pose"]
-    pos, quat, t_gt = gt["position"][:], gt["orientation"][:], gt["time"][:]
-    joint = f["unitree_hardware/joint_foot"]
-    q, dq, tau = joint["position"][:], joint["velocity"][:], joint["effort"][:]
-```
-
-To train or evaluate with the [SSIO code](https://github.com/SSIO-IMU/SSIO), download
-into `$DATA_ROOT/mipo_hdf5` and use the `MIPO_*` dataset configs.
+Unlike a plain NumPy dataset, the HDF5 files here are intentionally unsynchronized and
+uncalibrated: no per-topic resampling, no kalibr time-offset or extrinsic applied, no
+phone-to-robot clock correction. Reading the groups directly with `h5py` will silently skip all of
+that and produce a result that looks plausible but is misaligned. It is recommended to use, or
+faithfully reapply, the calibration from the [SSIO loader](https://github.com/SSIO-IMU/SSIO)
+(`Sequence.MIPO.MIPO_Sequence`, the `MIPO_*` dataset configs), which already implements it
+correctly. Download into `$DATA_ROOT/mipo_hdf5` and point the configs there.
 
 ## License and Attribution
 
 The recordings come from [Cerberus 2.0](https://github.com/ShuoYangRobotics/Cerberus2.0)
-(AGPL-3.0) by Shuo Yang, Zixin Zhang, Benjamin Bokser, and Zachary Manchester. Please also
-cite their work:
+(AGPL-3.0) by Shuo Yang, Zixin Zhang, Benjamin Bokser, and Zachary Manchester. Please also cite
+their work:
 
 ```bibtex
 @inproceedings{yang2023mipo,
@@ -121,7 +103,9 @@ cite their work:
 ```
 
 ## Citation
+
 https://ssio-imu.github.io
+
 ```bibtex
 @misc{qiu2026ssioselfsupervisedinertialodometry,
       title={SSIO: Self-Supervised Inertial Odometry},

download_mipo.py

@@ -0,0 +1,11 @@
+#!/usr/bin/env python3
+"""Download the MIPO dataset from Hugging Face."""
+
+from huggingface_hub import snapshot_download
+
+snapshot_download(
+    repo_id="Open-IO/mipo-go1-odometry",
+    repo_type="dataset",
+    revision="add-outdoor-data",  # drop this once the branch is merged to main
+    local_dir="mipo",
+)

indoor/README.md

@@ -0,0 +1,88 @@
+# MIPO: Indoor
+
+One HDF5 per source ROS bag. Recorded on a Unitree Go1 with OptiTrack motion-capture ground truth. See the [top-level README](../DATA_README.md) for original rosbag sourcing and attribution. This release covers the four indoor sessions from Cerberus2.0's MIPO indoor recordings listed below (`20230615_risqh_gait_analysis`, `20230620_risqh_robustness_test`, `230622_robustness`, `20230625_gait`). For ease of use, we converted this data to hdf5 format using the `Datasets/mipo_preprocess/MIPO_bag_to_hdf5.py` script. These HDF5s are what is on this hugging face repo as the rosbags are far larger but do not hold any more useful infromation for SSIO work.
+
+In the HDF5 files, each per-topic stream is stored at its native rate with its own ROS timestamp axis. Synchronization, resampling, filtering, and calibration are not applied.
+
+## Directory layout
+
+```
+indoor/
+├── mipo_time_offsets.json
+├── mipo_joint_stats.json
+├── 20230615_risqh_gait_analysis/
+│   └── *.h5
+├── 20230620_risqh_robustness_test/
+│   └── *.h5
+├── 20230625_gait/
+│   └── *.h5
+└── 230622_robustness/
+    └── *.h5
+```
+
+## HDF5 Groups
+
+See the [top-level README](../DATA_README.md#robot-sensor-data) for the shared robot-side schema
+(root attributes, body IMU, WT901 leg IMUs, joints + foot contact, and the joint name table),
+confirmed at the same rates on indoor files. What's specific to indoor:
+
+- `unitree_hardware/imu/orientation`: populated on most sequences (71/79). A few bags have
+  all-zero, and, as such, so do their HDF5s. Check before using.
+- `unitree_hardware/joint_foot/velocity`: Not used but special case where instead of zero beign on cols 12–15 like in all other data, the 5 `convexmpc` bags in `20230620_risqh_robustness_test/`, seem to have a 0/1 contact flag stored there instead.
+
+### `natnet_ros/Shuo_Go1/pose/` (OptiTrack GT pose, ~60 Hz, varies)
+
+| Dataset | Shape | Notes |
+|---|---|---|
+| `position` | (N, 3) | world-frame XYZ, meters |
+| `orientation` | (N, 4) | quaternion (x, y, z, w), world ← body |
+| `time` | (N,) | sec |
+
+No velocity stream.
+
+## Calibration
+
+MIPO sequences require two calibration corrections to align the MoCap (NatNet) ground truth with the IMU data: a **temporal offset** and a **spatial extrinsic (T_BS)**. Both are estimated per-sequence via kalibr and stored in `mipo_time_offsets.json`. These calibrations are applied in the SSIO MIPO Dataset Loader using the `mipo_time_offsets.json` file.
+
+**Time offset.** Convention: `t_imu = t_gt + time_offset_s`, applied by shifting GT timestamps so the GT and IMU streams land on a consistent resampled grid. The JSON stores a `status` field per sequence:
+- `"optimized"`: kalibr converged. Offset is a real measurement.
+- `"prior"`: kalibr did not converge. The stored value is the initial guess, not a meaningful estimate.
+
+Per-sequence `time_offset_s` is applied only when `status == "optimized"`. Otherwise the sequence falls back to the `_consensus` entry's `time_offset_s` (median over inlier sequences), the same consensus value described below for `T_BS`.
+
+**T_BS.** Transforms GT pose from the body frame (MoCap marker cluster) to the IMU sensor frame:
+
+    q_WS = q_WB ⊗ q_BS
+    p_WS = p_WB + R_WB @ t_BS
+
+Raw IMU data is untouched. Only GT orientation and position are rotated/translated. A single global T_BS constant is applied to every sequence rather than each sequence's own kalibr fit. The JSON does contain per-sequence T_BS estimates from kalibr, but these are not used (for a large fraction of sequences kalibr did not converge on T_BS (status=prior), yielding arbitrary rotations up to 159°). The global constant (the `_consensus` entry) is a consensus estimate from the well-calibrated inlier sequences and is more reliable.
+
+## Train/validation/test split exclusions
+
+All 79 sequences are published here. The pools actually used for model training are `MIPO_Train`
+(36), `MIPO_Valid` (12), and `MIPO_Test` (12), disjoint, 60 total
+(`MACIO_Pipeline2/Config/dataset/MIPO_{Train,Valid,Test}.yaml`). `MIPO_FullTest` is their union.
+
+Of the 19 sequences outside these pools, 6 are excluded for data-quality reasons worth noting
+here. 5 are GT mocap freeze/jump glitches, and remain in the self-supervised pseudo-label pool
+(`MIPO_Pseudo`) even though the ground truth itself is unusable there:
+
+| Sequence | Reason |
+|---|---|
+| `20230620_risqh_robustness_test/230620-risqh-trot-05-036-33square-kick3.h5` | GT freeze + jump (0.21 s, 0.26 m) |
+| `20230620_risqh_robustness_test/230620-risqh-trot-10-052-33square-convexmpc.h5` | GT freeze + jump (3.18 s, 0.93 m) |
+| `20230620_risqh_robustness_test/230620-risqh-trot-10-036-freewalk-convexmpc.h5` | GT freezes + jumps (3 × 1.2–1.5 s, 1.4–1.8 m) |
+| `20230620_risqh_robustness_test/230620-risqh-trot-10-048-33square-convexmpc.h5` | GT freeze + jump (0.66 s, 0.04 m) |
+| `20230620_risqh_robustness_test/230620-risqh-trot-10-044-33square-convexmpc.h5` | GT freeze + jump (3.42 s, 0.92 m) |
+
+The 6th, `20230625_gait/230625-risqh-trot-08-036-33square.h5`, has frozen IMU data at the end of
+the sequence and is excluded from every pool, including `MIPO_Pseudo`.
+
+The remaining 13 sequences are also outside the train/valid/test pools, excluded for reasons
+specific to this repo's own preprocessing pipeline rather than a defect in the recording itself.
+All 19 are still included as files in this release. Only pool membership differs.
+
+## Calibration Files
+
+- **`mipo_time_offsets.json`**: Per-sequence kalibr results, keyed by bag basename. Each entry has `time_offset_s`, `status`, and `T_BS` as described above, plus a `_consensus` entry holding the actual values applied (global `T_BS` and fallback for `time_offset_s`).
+- **`mipo_joint_stats.json`**: Dataset-wide normalization statistics (mean/std/count) for leg joint position, velocity, effort, and foot-contact effort, pooled across all indoor sequences.

indoor/mipo_joint_stats.json

@@ -1,22 +1,22 @@
 {
  "leg_pos": {
-  "mean": -0.34287641059613677,
-  "std": 1.1801247648104463,
+  "mean": -0.3428764045238495,
+  "std": 1.1801247596740723,
   "n": 4160532
  },
  "leg_vel": {
-  "mean": -0.00014636684348431014,
-  "std": 2.182366737941409,
+  "mean": -0.00014636688865721226,
+  "std": 2.1823668479919434,
   "n": 4160532
  },
  "leg_eff": {
-  "mean": 2.130135116730853,
-  "std": 5.217969268767851,
+  "mean": 2.1301350593566895,
+  "std": 5.2179694175720215,
   "n": 4160532
  },
  "contact_eff": {
-  "mean": 233.3156380113291,
-  "std": 288.7649325149753,
+  "mean": 233.3156280517578,
+  "std": 288.7649230957031,
   "n": 1386844
  }
 }

indoor/mipo_time_offsets.json

@@ -1344,7 +1344,6 @@
   },
   "_consensus": {
     "description": "Median of inlier optimized calibrations (euler < 10 deg, trans < 0.3m, time_offset < 0.5s)",
-    "time_offset_s": -0.11408568477439143,
     "n_total": 78,
     "n_optimized": 47,
     "n_inliers": 40,
@@ -1409,4 +1408,4 @@
       "230625-risqh-trot-06-040-8figure"
     ]
   }
-}
+}

outdoor/CMU garage/230828-cmu-trot-06-040-east-campus-garage-bad-gps.h5

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f844581e97b6f43618607bc4c722a65ba60505d48c5e5a07c39d4b9da9baf154
+size 60131765

outdoor/README.md

@@ -0,0 +1,149 @@
+# MIPO: Outdoor
+
+One HDF5 per source ROS bag, plus a densified position track per sequence. Five outdoor trials in four Pittsburgh locations (outdoor trails, track, and a campus garage + paved walk), recorded on a Unitree Go1 with no motion capture. The only position source is a phone (GPS + IMU) that was **carried by the operator walking with the robot**, not mounted on it. See the [top-level README](../DATA_README.md) for original data sourcing and attribution.
+
+In the HDF5 files, each per-topic stream is stored at its native rate with its own timestamp axis Synchronization, resampling, filtering, and calibration are not applied.
+
+## Visual Reference, NO Ground Truth
+
+There is **no robot ground truth outdoors.** The phone records the operator's path, which is typically a few meters behind or beside the robot. On all five sequences, the phone IMU shows human step and stride frequencies and no robot trot frequency. Tts angular rate is also inconsistent with a rigid mount on the robot. Everything position-related in `outdoor/` describes the operator's path. Use it as a georeferenced visual reference for the shape of the robot's route, not to measure velocity or orientation error. It can also be a coarse translation GT. The onboard AHRS orientation is the only robot attitude reference that comes from the Go1's onboard AHRS filter. Even that is a weak ground truth. Though it is a significant improvement over IMU angular velocity integration.
+
+Phone GPS is also coarse. Measurements arrive at ~1 Hz. Latitude and longitude are stored to only 4 decimal places, which is a grid of about 1.1 m north by 0.85 m east. The operator walked at only ~0.6 to 0.8 m/s, so motion between consecutive measurements is often smaller than one grid step. Any track fitted to these measurements looks like a low-passed GPS trace. Expect errors of a meter or more on top of the operator-to-robot offset. Nonetheless, a modified version of our PGO was run to fust the phone's IMU data with the GPS measurments to smooth and interpolate the reference trajectory. This is what is stored in the `pgo_densified_ref` directory.
+
+## How the outdoor HDF5s were produced
+
+1. **Robot bags to HDF5.** Same conversion as indoor: native rates, robot laptop clock, no resampling or calibration.
+2. **Phone recordings.** From the original release, each sequence has a MATLAB Mobile `.mat` recording with phone IMU, magnetometer, attitude and GPS, decoded to plain arrays. The recorded timestamps carried no time zone, so they were interpreted as US Eastern time, which gives true UTC POSIX times. The receiver's Doppler ground speed, course over ground and horizontal accuracy were kept alongside latitude, longitude and altitude.
+3. **Merge.** The decoded phone data is stored as a `/phone` group in each sequence's HDF5. Robot streams are copied unchanged.
+4. **Clock offset.** The phone and robot clocks differ by a few seconds. The offset was estimated per sequence by cross-correlating the GPS trajectory turn rate with the onboard AHRS turn rate. This uses heading only, so it does not depend on how the phone was held. Stored times stay on their native clocks, and the offsets are released as the `ahrs_gps_yaw_offsets.json` manifest.
+
+## How the densified reference was produced
+
+GPS alone is too sparse and discretized for a smooth path. For each sequence, a physically-consistent, smoothed, and interpolated trajectory was fit to the phone data with a pose-graph optimization:
+
+- **States.** Each keyframe (every 0.4 to 1 s) has a phone attitude, velocity and position in a local ENU frame centered on the first GPS fix.
+- **Motion between keyframes.** Phone IMU readings are preintegrated between keyframes forming a binary constraint, with gravity modeled explicitly against each keyframe's attitude.
+- **Position.** Every GPS fix is a unary position constraint, and all fixes are used.
+- **Stops.** Detected stationary periods add zero-velocity constraints.
+- **Start.** The first keyframe is placed at the first GPS fix, with its initial velocity from the GPS Doppler speed and course.
+
+| Sequence | Keyframe rate | Phone IMU used | RMS distance to GPS fixes | Result (qualitative) |
+|---|---|---|---|---|
+| schedley park | 1.0 Hz | 100 Hz | 0.55 m | Good |
+| mill19 trail | 2.5 Hz | 100 Hz | 0.45 m | Good |
+| st mary catholic cemetery | ~2 Hz | 25 Hz (decimated) | 0.51 m | Good |
+| frick park | 2.0 Hz | 10 Hz (native) | 1.08 m | Decent |
+| CMU garage | ~2 Hz | 25 Hz (decimated) | n/a, see below | Poor |
+
+The RMS distance only measures how closely the track follows the GPS fixes. It is not an accuracy metric because the fixes themselves are coarse and record the operator's trajectory, not the robot. As such, the PGO parameters were tuned to qualitatively match smooth and phyiscally reasonable trajectories (no large excursions, odd loops, and generally stable behavior). Qualitative results: 
+- **Good.** The track follows the GPS cleanly.
+- **Decent (frick park).** The phone IMU was recorded at only 10 Hz. Small attitude errors between fixes leak gravity into the motion estimate, which leaves small excursions that no keyframe setting removed.
+- **Poor (CMU garage).** CMU garage, even though salvaged as best as we could, had to be split into outdoor pieces on either side of the garage with individual PGO solves. The pieces were then sanitized for odd beginning and end-of-outdoor-period effects where GPS was starting to become unreliable. Overall, the positions are less reliable, and the CMU garage sequence has a large gap (~282 s with no position) in and around the garage. See its dedicated `seg`/`break_kf` keys below.
+
+Sequences with decimated phone IMU (CMU garage and st mary catholic cemetery) were too large to have their full-rate, 100 Hz IMU as part of one solve on our local machine. Tests of the solve on partial trajectories showed that they decmiation was not detrimental to the solves.
+
+## Dropped sequences
+
+The original MIPO dataset's Wightman Park sequences were dropped for lack of phone GPS data and AHRS filter orientation.
+
+## Directory layout
+
+```
+outdoor/
+├── ahrs_gps_yaw_offsets.json
+├── CMU garage/
+│   └── 230828-cmu-trot-06-040-east-campus-garage-bad-gps.h5
+├── frick park/
+│   └── 230626-frick-park-trot-06-04-arch.h5
+├── mill19 trail/
+│   └── 230628-mil19-trot-07-039-wild1445.h5
+├── schedley park/
+│   └── 230630-schen-trot-08-038-loop-tennis.h5
+├── st mary catholic cemetery/
+│   └── 230828-st-mary-catholic-cemetery-trot-06-040-loop.h5
+└── pgo_densified_ref/
+    └── <place>/<seq>.npz        (same relative path as the .h5)
+```
+
+## Sequences
+
+| Place | Sequence | Robot duration | Phone IMU | GPS fixes | Densified ref quality |
+|---|---|---|---|---|---|
+| schedley park | `230630-schen-trot-08-038-loop-tennis` | 299 s | 100 Hz | 291 | Good |
+| mill19 trail | `230628-mil19-trot-07-039-wild1445` | 410 s | 100 Hz | 410 | Good |
+| st mary catholic cemetery | `230828-st-mary-catholic-cemetery-trot-06-040-loop` | 707 s | 100 Hz | 698 | Good |
+| frick park | `230626-frick-park-trot-06-04-arch` | 366 s | **10 Hz** | 348 | Decent |
+| CMU garage | `230828-cmu-trot-06-040-east-campus-garage-bad-gps` | 645 s | 100 Hz | 476 | Poor (including large gap) |
+
+## HDF5 Groups
+
+See the [top-level README](../DATA_README.md#robot-sensor-data) for the shared robot-side schema (root attributes, body IMU, WT901 leg IMUs, joints + foot contact, and the joint name table). Outdoor has no `gt_topic` attribute, because there is no robot ground-truth topic outdoors. What's specific to outdoor:
+
+### `phone/` (operator-carried phone)
+
+The phone streams were recorded with MATLAB Mobile. All `time` arrays are on the **phone clock** (POSIX seconds UTC, float64), which is offset from the robot clock by a few seconds (see [Calibration](#calibration)).
+
+| Dataset | Shape | Units / notes |
+|---|---|---|
+| `Acceleration/xyz` | (N, 3) | m/s², phone body frame, gravity included |
+| `AngularVelocity/xyz` | (N, 3) | rad/s, phone body frame |
+| `MagneticField/xyz` | (N, 3) | µT, phone body frame |
+| `Orientation/xyz` | (N, 3) | deg, the phone's own fused attitude (iOS Core Motion). Euler angles stored as [X, Y, Z], ZYX order, frame rotation from the phone world frame (gravity-aligned, x ≈ magnetic north) to the phone body. |
+| `*/time` | (N,) | sec, one per stream above |
+| `Position/time` | (M,) | sec, ~1 Hz |
+| `Position/lat`, `Position/lon` | (M,) | deg, WGS84 (stored to 4 decimal places) |
+| `Position/alt` | (M,) | m |
+| `Position/speed` | (M,) | m/s, GPS Doppler ground speed |
+| `Position/course` | (M,) | deg clockwise from true north, GPS course over ground |
+| `Position/hacc` | (M,) | m, receiver-reported horizontal accuracy (median ~4.7 m) |
+
+Phone IMU streams are ~100 Hz on every sequence except frick park, which is 10 Hz.
+
+`phone` group attributes:
+
+| Attribute | Meaning |
+|---|---|
+| `refloc` | `[lat, lon, alt]` of the first GPS fix. This is the origin of the ENU frame used by `pgo_densified_ref/`. |
+| `refposixtime` | POSIX reference time of the recording (phone clock). |
+| `source_mat` | Filename of the original phone `.mat` recording. |
+| `phone_time_shift_applied_s` | Always 0. No clock shift is baked into the stored times. |
+
+## `pgo_densified_ref/<place>/<seq>.npz` (densified operator path)
+
+A smooth position track per sequence, fitted to the phone IMU and GPS. **It is the operator's phone path, not robot ground truth.** Use it as a georeferenced visual reference or coarse translation GT only. Velocity or orientation error metrics against it are not meaningful.
+
+- **World frame:** local-tangent ENU (x east, y north, z up, meters). The origin is `phone.attrs["refloc"]` of the matching HDF5. The CMU garage file stores its own `refloc`, so use that one for CMU.
+- **Clock:** all times are on the **phone clock**. Add the sequence's offset (see [Calibration](#calibration)) to place them on the robot clock.
+
+| Key | Shape | Meaning |
+|---|---|---|
+| `kf_t` | (K,) | Keyframe times, sec, phone clock. 1 to 2.5 Hz depending on the sequence. |
+| `pos` | (K, 3) | Keyframe position, ENU, m. |
+| `vel_world` | (K, 3) | Keyframe velocity, ENU, m/s. |
+| `rot` | (K, 4) | Keyframe phone attitude, quaternion (x, y, z, w), ENU ← phone body. |
+| `gps_t` | (M,) | Times of the GPS fixes used, sec, phone clock. |
+| `gps_enu` | (M, 3) | Those fixes in the same ENU frame, m. |
+| `speed` | (M,) | GPS Doppler speed at those fixes, m/s. |
+| `holdout_idx` | (0,) | Always empty, because every GPS fix was used. Absent from the CMU file. |
+
+CMU garage only:
+
+| Key | Shape | Meaning |
+|---|---|---|
+| `seg` | (K,) | Segment id per keyframe. 0 = before the garage (~29 s), 1 = after it (~329 s). |
+| `break_kf` | () | Index of the first keyframe of segment 1. The ~282 s between the segments has no position estimate. Do not interpolate across it. |
+| `refloc` | (3,) | `[lat, lon, alt]` ENU origin for this file. |
+| `seq_name` | () | Sequence key, `<place>/<seq>` (HDF5 path relative to `outdoor/` without `.h5`). |
+
+## Calibration
+
+The raw streams are untouched. Two corrections are applied when the MIPO Dataset loader.
+
+**Phone-to-robot clock offset.** The phone and robot clocks differ by a few seconds. Convention: `t_robot = t_phone + offset`. This is applied to the phone GPS and densified-reference timestamps only. Per sequence, use `offset_begin_s` when `begin_reliable` is true, otherwise `offset_s`. A sequence with `trusted: false` has no usable offset, so its GPS is not placed on the robot clock. Only CMU garage is untrusted, because its GPS outage makes the estimate ambiguous. The trusted offsets range from −4.4 s to −0.3 s. The operator follows the robot, so the offset also absorbs some following lag. Treat it as accurate to about a second.
+
+**AHRS tilt.** Against indoor motion capture on still windows, the onboard AHRS shows a constant tilt bias. Its orientation is corrected in the body frame as `R = R_ahrs · Exp([0.795°, 0.131°, 0])` (roll, pitch, yaw). The onboard AHRS is the only robot attitude reference outdoors.
+
+## Manifests
+
+- **`ahrs_gps_yaw_offsets.json`**: Per-sequence phone-to-robot clock offsets, keyed by HDF5 path relative to `outdoor/` without `.h5`. Despite the name, these are time offsets, not yaw offsets (were measured by cross-correlating GPS yaw rate with onboard AHRS yaw rate). Each entry has `offset_s`, `offset_begin_s`, `begin_reliable` and `trusted` as described above. The rest are diagnostics: `peak_corr`, `peak_corr_begin`, `n_turn_samples`, `stability_spread_s`, `drift_first_to_last_s`, `drift_slope_s_per_min`, `windows` (90 s sliding-window estimates) and `note`.
+- **Joint normalization.** Joints are standardized with `indoor/mipo_joint_stats.json` from indoor data split since the normalization data comes from the train split.

outdoor/ahrs_gps_yaw_offsets.json

@@ -0,0 +1,315 @@
+{
+  "convention": "offset_s added to phone GPS timestamps yields the bag/robot clock time. Estimated by cross-correlating GPS course-over-ground turn-rate vs onboard AHRS yaw-rate (model-independent, data-only).",
+  "method": "GPS lat/lon->ENU, smooth+multi-second-baseline course-over-ground, mask <0.5 m/s, unwrap+differentiate to turn-rate; AHRS quaternion yaw (body x-axis in horizontal plane), decimate, unwrap+differentiate to yaw-rate; resample both to 10 Hz, 3 s gap-aware smoothing, normalized cross-correlation over +/-30 s with parabolic sub-sample peak; offset_s=peak_lag such that bag=phone+offset_s. offset_begin_s uses only the first ~120 s (operator closest to robot, least accumulated AHRS drift) and is the cleaner start-time clock offset; windows[] is a 90 s/50%-overlap sliding-window diagnostic for time-varying following-lag drift. The full-sequence estimate searches +/-30 s; the windowed and beginning estimates search a narrow +/-8 s band (short windows have few turns and a wide search aliases onto spurious peaks; the true offset is known to be small from the stable full-sequence value).",
+  "sequences": {
+    "CMU garage/230828-cmu-trot-06-040-east-campus-garage-bad-gps": {
+      "offset_s": 15.46,
+      "offset_begin_s": -5.28,
+      "begin_reliable": false,
+      "peak_corr": 0.2367,
+      "peak_corr_begin": 0.1914,
+      "n_turn_samples": 3991,
+      "stability_spread_s": 27.048,
+      "trusted": false,
+      "windows": [
+        {
+          "win_start_s": 0.0,
+          "offset_s": -5.49,
+          "peak_corr": 0.2339
+        },
+        {
+          "win_start_s": 45.0,
+          "offset_s": -4.135,
+          "peak_corr": 0.0592
+        },
+        {
+          "win_start_s": 90.0,
+          "offset_s": -0.906,
+          "peak_corr": 0.8778
+        },
+        {
+          "win_start_s": 135.0,
+          "offset_s": null,
+          "peak_corr": null
+        },
+        {
+          "win_start_s": 180.0,
+          "offset_s": null,
+          "peak_corr": null
+        },
+        {
+          "win_start_s": 225.0,
+          "offset_s": -8.0,
+          "peak_corr": -0.066
+        },
+        {
+          "win_start_s": 270.0,
+          "offset_s": 0.839,
+          "peak_corr": 0.2888
+        },
+        {
+          "win_start_s": 315.0,
+          "offset_s": 1.623,
+          "peak_corr": 0.5767
+        },
+        {
+          "win_start_s": 360.0,
+          "offset_s": -8.0,
+          "peak_corr": 0.0962
+        },
+        {
+          "win_start_s": 405.0,
+          "offset_s": -8.0,
+          "peak_corr": 0.0351
+        },
+        {
+          "win_start_s": 450.0,
+          "offset_s": 2.858,
+          "peak_corr": 0.26
+        },
+        {
+          "win_start_s": 495.0,
+          "offset_s": -2.534,
+          "peak_corr": 0.2475
+        },
+        {
+          "win_start_s": 540.0,
+          "offset_s": -0.813,
+          "peak_corr": 0.6385
+        }
+      ],
+      "drift_first_to_last_s": 0.094,
+      "drift_slope_s_per_min": 0.152,
+      "note": "UNTRUSTED: ~145 s GPS outage; cross-correlation multi-modal. Do not use."
+    },
+    "frick park/230626-frick-park-trot-06-04-arch": {
+      "offset_s": -2.317,
+      "offset_begin_s": -2.193,
+      "begin_reliable": true,
+      "peak_corr": 0.4753,
+      "peak_corr_begin": 0.5457,
+      "n_turn_samples": 2695,
+      "stability_spread_s": 0.487,
+      "trusted": true,
+      "windows": [
+        {
+          "win_start_s": 0.0,
+          "offset_s": -1.836,
+          "peak_corr": 0.6873
+        },
+        {
+          "win_start_s": 45.0,
+          "offset_s": -1.814,
+          "peak_corr": 0.5161
+        },
+        {
+          "win_start_s": 90.0,
+          "offset_s": 3.387,
+          "peak_corr": 0.303
+        },
+        {
+          "win_start_s": 135.0,
+          "offset_s": -2.644,
+          "peak_corr": 0.374
+        },
+        {
+          "win_start_s": 180.0,
+          "offset_s": -2.414,
+          "peak_corr": 0.688
+        },
+        {
+          "win_start_s": 225.0,
+          "offset_s": -2.361,
+          "peak_corr": 0.6149
+        },
+        {
+          "win_start_s": 270.0,
+          "offset_s": -4.756,
+          "peak_corr": 0.392
+        }
+      ],
+      "drift_first_to_last_s": -2.92,
+      "drift_slope_s_per_min": -0.745,
+      "note": "moderate correlation - offset stable, treat with mild caution"
+    },
+    "mill19 trail/230628-mil19-trot-07-039-wild1445": {
+      "offset_s": -1.566,
+      "offset_begin_s": -0.342,
+      "begin_reliable": true,
+      "peak_corr": 0.4424,
+      "peak_corr_begin": 0.3787,
+      "n_turn_samples": 3095,
+      "stability_spread_s": 0.403,
+      "trusted": true,
+      "windows": [
+        {
+          "win_start_s": 0.0,
+          "offset_s": -0.521,
+          "peak_corr": 0.4419
+        },
+        {
+          "win_start_s": 45.0,
+          "offset_s": 5.853,
+          "peak_corr": 0.2131
+        },
+        {
+          "win_start_s": 90.0,
+          "offset_s": -6.078,
+          "peak_corr": 0.2558
+        },
+        {
+          "win_start_s": 135.0,
+          "offset_s": 2.164,
+          "peak_corr": 0.284
+        },
+        {
+          "win_start_s": 180.0,
+          "offset_s": -1.773,
+          "peak_corr": 0.4539
+        },
+        {
+          "win_start_s": 225.0,
+          "offset_s": -1.056,
+          "peak_corr": 0.6495
+        },
+        {
+          "win_start_s": 270.0,
+          "offset_s": -1.738,
+          "peak_corr": 0.8123
+        },
+        {
+          "win_start_s": 315.0,
+          "offset_s": -1.822,
+          "peak_corr": 0.6788
+        }
+      ],
+      "drift_first_to_last_s": -1.301,
+      "drift_slope_s_per_min": 0.02,
+      "note": "moderate correlation - offset stable, treat with mild caution"
+    },
+    "schedley park/230630-schen-trot-08-038-loop-tennis": {
+      "offset_s": -1.609,
+      "offset_begin_s": -0.315,
+      "begin_reliable": true,
+      "peak_corr": 0.5013,
+      "peak_corr_begin": 0.469,
+      "n_turn_samples": 2772,
+      "stability_spread_s": 0.409,
+      "trusted": true,
+      "windows": [
+        {
+          "win_start_s": 0.0,
+          "offset_s": -0.153,
+          "peak_corr": 0.5251
+        },
+        {
+          "win_start_s": 45.0,
+          "offset_s": -1.911,
+          "peak_corr": 0.3185
+        },
+        {
+          "win_start_s": 90.0,
+          "offset_s": -2.376,
+          "peak_corr": 0.6159
+        },
+        {
+          "win_start_s": 135.0,
+          "offset_s": -2.802,
+          "peak_corr": 0.6503
+        },
+        {
+          "win_start_s": 180.0,
+          "offset_s": -1.651,
+          "peak_corr": 0.593
+        }
+      ],
+      "drift_first_to_last_s": -1.498,
+      "drift_slope_s_per_min": -0.518
+    },
+    "st mary catholic cemetery/230828-st-mary-catholic-cemetery-trot-06-040-loop": {
+      "offset_s": -4.371,
+      "offset_begin_s": -8.0,
+      "begin_reliable": false,
+      "peak_corr": 0.2125,
+      "peak_corr_begin": 0.1566,
+      "n_turn_samples": 4195,
+      "stability_spread_s": 0.423,
+      "trusted": true,
+      "windows": [
+        {
+          "win_start_s": 0.0,
+          "offset_s": -8.0,
+          "peak_corr": 0.1941
+        },
+        {
+          "win_start_s": 45.0,
+          "offset_s": -8.0,
+          "peak_corr": 0.262
+        },
+        {
+          "win_start_s": 90.0,
+          "offset_s": -5.271,
+          "peak_corr": 0.239
+        },
+        {
+          "win_start_s": 135.0,
+          "offset_s": 8.0,
+          "peak_corr": 0.1996
+        },
+        {
+          "win_start_s": 180.0,
+          "offset_s": -3.502,
+          "peak_corr": 0.3794
+        },
+        {
+          "win_start_s": 225.0,
+          "offset_s": -3.9,
+          "peak_corr": 0.48
+        },
+        {
+          "win_start_s": 270.0,
+          "offset_s": -1.107,
+          "peak_corr": 0.2102
+        },
+        {
+          "win_start_s": 315.0,
+          "offset_s": -4.14,
+          "peak_corr": 0.3831
+        },
+        {
+          "win_start_s": 360.0,
+          "offset_s": -3.516,
+          "peak_corr": 0.7283
+        },
+        {
+          "win_start_s": 405.0,
+          "offset_s": -3.73,
+          "peak_corr": 0.5263
+        },
+        {
+          "win_start_s": 450.0,
+          "offset_s": -4.025,
+          "peak_corr": 0.2633
+        },
+        {
+          "win_start_s": 495.0,
+          "offset_s": 6.682,
+          "peak_corr": 0.3619
+        },
+        {
+          "win_start_s": 540.0,
+          "offset_s": 7.196,
+          "peak_corr": 0.5497
+        },
+        {
+          "win_start_s": 585.0,
+          "offset_s": -5.125,
+          "peak_corr": 0.3901
+        }
+      ],
+      "drift_first_to_last_s": 2.875,
+      "drift_slope_s_per_min": 0.998,
+      "note": "weak: low full-sequence peak correlation (0.21) though offset is stable across windows - many gentle turns, limited SNR; beginning-only estimate UNRELIABLE (begin_corr=0.16, pinned at search rail) - use full-sequence offset_s instead"
+    }
+  }
+}

outdoor/frick park/230626-frick-park-trot-06-04-arch.h5

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63a3b42ae65c05553c10bd43fbae1bd44616dfb9b1da2c7b9085dad2549b7819
+size 30037644

outdoor/mill19 trail/230628-mil19-trot-07-039-wild1445.h5

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5fe621cb90eb4d67fa89270ddef10f43fb74de9a71b349eeb35eb54a69e5af4
+size 39010583

outdoor/pgo_densified_ref/CMU garage/230828-cmu-trot-06-040-east-campus-garage-bad-gps.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dc7d160fd2bfafb8c56d36ed75e643585eff1280ac5a27be20f1362d6c3e15a
+size 91720

outdoor/pgo_densified_ref/frick park/230626-frick-park-trot-06-04-arch.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9276daf2c95a47c18ec621d9999e5d85b422c65282a6fa6975225f0978064de2
+size 76964

outdoor/pgo_densified_ref/mill19 trail/230628-mil19-trot-07-039-wild1445.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad512da528148e5dbca62b1f48341de4833f6c0088dda7ba478ead939e86b59f
+size 108484

outdoor/pgo_densified_ref/schedley park/230630-schen-trot-08-038-loop-tennis.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1e80c322dd3cec39b649f6b1bdcc0f6e7e6fababe72bd78d8e369cdb7c9d4f2
+size 39220

outdoor/pgo_densified_ref/st mary catholic cemetery/230828-st-mary-catholic-cemetery-trot-06-040-loop.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f504ed2d57bd884c0b1775e4f2887f1c4976cf632bdeefe375d902303c98d4f
+size 157580

outdoor/schedley park/230630-schen-trot-08-038-loop-tennis.h5

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:941b4af7bf70db54730daaa8b6ff29aa63d69ba7d0a7f8806c47cfcd2ded7645
+size 28098630

outdoor/st mary catholic cemetery/230828-st-mary-catholic-cemetery-trot-06-040-loop.h5

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c633922300176b72af89ca5bbd485caadd4a9742e16d0a956e88e6c807fa7cf
+size 66087559