task2_visual_imu/README.md

Task 2 — Visual + IMU

Purpose

Synchronized visual-inertial clips for testing joint reasoning over RGB frames, trajectory, and raw gyroscope/accelerometer signals.

Files included

task2_visual_imu/
├── clips/
│   ├── clip_001/ .. clip_007/
│   │   ├── frames/F01.png .. F10.png
│   │   ├── trajectory_2d.png
│   │   ├── imu.csv
│   │   ├── gyro_plot.png
│   │   ├── accel_plot.png
│   │   ├── frame_imu_alignment.json
│   │   └── metadata.json
│   └── ...
├── manifest.jsonl
└── README.md

7 clips (target was 8-12; only 7 of the 8 Stage A source windows have real IMU — S3, TUM RGB-D walking_xyz, is RGB-D-only with no inertial sensor, and is correctly excluded, not padded with fabricated IMU data).

clip_id	source	sequence	motion_type
clip_001	S1	EuRoC MH_04_difficult	turn
clip_002	S2	EuRoC V1_03_difficult	turn
clip_003	S4	TUM-VI corridor1_512_16	corridor
clip_004	S5	TUM-VI corridor4_512_16	straight
clip_005	S6	TUM-VI magistrale1_512_16	room
clip_006	S7	TUM-VI room1_512_16	mixed
clip_007	S8	TUM-VI room2_512_16	rotation

How they were generated

Time windows and frame sampling are the exact same ones used for the Stage A benchmark's healthy source clips (configs/extraction_config.yaml at https://github.com/deansmile/agentic-slam), reused here for consistency — frames/ and trajectory_2d.png are copied from task3_open_weight_vlm/cases/<Sx>_healthy/ in this same upload (already verified to be the exact frames the original harness used).

imu.csv is extracted directly from each source dataset's raw mav0/imu0/data.csv (EuRoC/TUM-VI native format), filtered to the same absolute time window as the frames, with timestamps re-based to be relative to the clip's start (0.0 = first frame).

imu.csv schema

timestamp_sec,gx,gy,gz,ax,ay,az

• timestamp_sec: seconds, relative to clip start (not absolute unix time — see metadata.json's clip_start_timestamp_abs to recover absolute time if needed).
• gx,gy,gz: angular velocity, rad/s (gyroscope).
• ax,ay,az: linear acceleration, m/s² (accelerometer, includes gravity — not gravity-compensated, matching the raw sensor convention).
• Original units and axes are preserved exactly as recorded by the EuRoC/TUM-VI IMU (ADIS16448-class MEMS IMU, nominal 200 Hz) — not rescaled. Nominal rate is documented in each dataset's own mav0/imu0/sensor.yaml (EuRoC) where present; TUM-VI ships the same column format without a separate sensor.yaml, same nominal 200 Hz per the TUM-VI paper. metadata.json's n_imu_samples gives the actual observed sample count per clip (~2790-2800 for each 14 s window here, consistent with ~200 Hz).

gyro_plot.png and accel_plot.png are generated directly from the exact imu.csv shipped alongside them (not from a separate/different source) — regenerating the plots from imu.csv should reproduce them exactly.

frame_imu_alignment.json explicitly maps every F01-F10 frame to its nearest IMU sample (by timestamp): frame_id, absolute and clip-relative frame timestamp, and the matched IMU row's index and relative timestamp.

How to use them with the GitHub code

Frames/trajectory use the same convention as trajectory_failure_benchmark/ at https://github.com/deansmile/agentic-slam. imu.csv's schema is intentionally simple (flat CSV, documented units) so it can be loaded directly with pandas.read_csv / numpy.loadtxt without any project-specific code.

Which files are model-visible

frames/, trajectory_2d.png, imu.csv (or gyro_plot.png/ accel_plot.png if you want a visual IMU representation instead of raw numbers) are all model-visible.

Which files contain hidden labels

None of these clips have hidden failure labels — this task is about joint visual+inertial reasoning on healthy motion, not failure detection (contrast with Task 1/3, which do carry hidden labels in their metadata.json/benchmark_manifest.jsonl). metadata.json here is pure provenance (source sequence, config, timestamps, units), safe to inspect but not meaningful as a "ground truth label" in the failure-detection sense.

Known limitations

• Only 7 clips (not 8-12) because only 7 of the 8 Stage A sources have real IMU; S3 (TUM RGB-D) is excluded rather than backfilled with synthetic/unrelated IMU data.
• All 7 clips reuse Stage A's _healthy condition windows (no visual-inertial failure cases are included here — Task 2 is about joint visual+IMU reasoning capability, not failure detection).
• Accelerometer readings include gravity (not gravity-compensated); if your use case needs gravity removed, that requires the sequence's known initial orientation, which is not provided here.

Source / provenance

• Time windows: /scratch/ds5725/agentic_slam_github/trajectory_failure_benchmark/configs/extraction_config.yaml
• Frames/trajectory: task3_open_weight_vlm/cases/<Sx>_healthy/ (this upload)
• Raw IMU: /scratch/ds5725/agentic_slam/datasets/{euroc,tum_vi}/<sequence>/mav0/imu0/data.csv

Minimal example workflow

import json
import pandas as pd

imu = pd.read_csv("clips/clip_006/imu.csv")  # S7, TUM-VI room1
align = json.load(open("clips/clip_006/frame_imu_alignment.json"))

f05 = next(f for f in align["frames"] if f["frame_id"] == "F05")
nearby_imu = imu.iloc[max(0, f05["nearest_imu_row_index"] - 5):f05["nearest_imu_row_index"] + 5]
print(nearby_imu)