Support Information
Hardware:
- Aircraft: ModalAI Starling 2 Max (C29)
- SKU:
MRB-D0012-4-V4-C29-T8-M36 - VOXL platform: VOXL2 M0054
- Tracking cameras:
tracking_frontandtracking_down(AR0144) - IMU: ICM42688
Software version:
- System image:
1.8.08-M0054-14.1a-perf - VOXL suite:
1.6.4~beta5 - OpenVINS:
voxl-open-vins-server 0.6.1
These are the software versions recorded during our investigation.
Configuration / services:
- Two-camera OpenVINS configuration using
tracking_frontandtracking_down. - Official configuration preset:
starling2_2cam. - IMU service:
voxl-imu-server. - VIO service:
voxl-open-vins-server. - OpenVINS was recently restored using
voxl-configure-open-vins starling2_2cam. - Previous diagnostic tests included changes to online camera-extrinsic calibration settings.
- A startup-order issue between OpenVINS and the IMU service was identified.
What I am trying to do:
We are developing autonomous indoor navigation using the Starling 2 Max and require reliable, repeatable metric odometry.
Our environment is an ordinary indoor workshop containing shelving, storage racks, tables, and workbenches. There is nothing particularly unusual about the environment.
We also intend to purchase a second identical Starling 2 Max, while retaining our existing aircraft.
We would like the new unit to arrive with the most recent complete software configuration officially validated for this specific aircraft, including any available improvements relevant to OpenVINS reliability.
What actually happens:
OpenVINS produces substantially different metric-scale estimates between diagnostic sessions.
During approximately 1.4-meter vertical movements, we measured estimated displacements of:
- 1.41 m: approximately correct.
- 0.32 m: approximately 23% of the physical displacement.
- 0.46 m: approximately 33% of the physical displacement.
Horizontal odometry also exhibited inconsistent scale.
Additional observations:
- OpenVINS occasionally initializes before the IMU service is ready.
- IMU propagation-window configuration differs depending on startup conditions.
- Timestamp regressions of approximately 33.3 ms have occurred.
BAD_TIMESTAMPerrors have triggered OpenVINS automatic resets.- Visual-tracking quality and trajectory closure errors vary considerably between test sessions.
The root cause of the metric-scale inconsistency remains undetermined.
Steps to reproduce:
- Start the aircraft and initialize the two-camera OpenVINS system.
- Confirm that the IMU and tracking cameras are operating.
- Manually carry the aircraft through an approximately 1.4-meter vertical movement.
- Execute an approximately 2 m × 2 m horizontal square trajectory.
- Return toward the starting position.
- Compare OpenVINS displacement against the approximate physical movement.
- Repeat under another startup or initialization condition.
These were manually executed diagnostic movements without motor actuation. Some sessions involved configuration changes and are not identical controlled repetitions.
Logs / pictures:
No files are attached to this post.
Relevant errors observed during our investigation:
WARNING: could not read IMU pipe info JSON for 'imu_apps'
[HEALTH] ERROR: Sensor measurements with bad timestamps
AUTO-RESET RECOMMENDED Error code(s) detected 0x8000
The measured results and technical observations are summarized in the following report.
Detailed Technical Context
1. Background and purpose
Hello ModalAI Engineering Team,
We currently own a Starling 2 Max (C29), which we use for autonomous indoor robotics development.
We are planning to purchase a second identical Starling 2 Max and would appreciate your engineering recommendations regarding the most appropriate software configuration for the new aircraft.
We intend to keep our existing drone. We are not requesting a replacement or exchange.
Our objective is to continue using the existing aircraft for development while acquiring an additional unit with the most recent, officially validated software available for this particular hardware.
The reason for our inquiry is that we have encountered significant inconsistencies in OpenVINS metric-scale estimation.
We have investigated multiple possible contributing factors, including initialization behavior, camera–IMU synchronization, visual-tracking quality, calibration settings, and timestamp ordering.
We have not conclusively determined the cause.
However, several recent ModalAI forum discussions and OpenVINS software developments appear relevant to our observations.
We would appreciate your engineering assessment and clarification on whether the newer software configurations address any of these behaviors.
2. Operating environment
Our drone operates in an ordinary indoor workshop.
The environment contains shelves, storage racks, tables, workbenches, and other typical stationary workshop objects.
It is not a specialized motion-capture laboratory or an unusual visual environment.
We understand that visual-inertial odometry performance depends on factors such as lighting, feature availability, camera exposure, motion, synchronization, and calibration.
We are not claiming that all environmental or visual-tracking factors have been eliminated.
However, our primary concern is not ordinary odometry drift.
Our concern is the significant variation in metric scale between test sessions.
3. Observed metric-scale inconsistency
We performed several manually executed experiments with the aircraft's motors inactive.
A typical test involved an approximately measured vertical ascent of 1.4 meters, followed by a horizontal square trajectory with approximately 2-meter sides.
The objective was to evaluate how accurately OpenVINS represented the physical movement.
Representative vertical measurements included:
| Test | Approximate physical movement | OpenVINS displacement | Scale |
|---|---|---|---|
| A | 1.40 m | 1.41 m | 1.01× |
| B | 1.40 m | 0.32 m | 0.23× |
| C | 1.40 m | 0.46 m | 0.33× |
We also observed inconsistent horizontal-scale estimates during approximately 2-meter movements.
Some sessions produced estimates relatively close to the physical trajectory, while others substantially underestimated displacement.
We recognize that these were hand-carried tests rather than motion-capture measurements.
They also represent different diagnostic sessions, some of which involved changes to the software configuration.
Nevertheless, the observed differences are large enough to raise concerns about initialization consistency or another aspect of the VIO estimation process.
We do not know whether the cause is software initialization, camera–IMU calibration, timestamp handling, visual tracking, or another factor.
This is the central technical question we would like ModalAI engineering to address.
4. IMU and OpenVINS startup behavior
During our investigation, we identified an issue involving the startup order of OpenVINS and the IMU service.
On some starts, OpenVINS initialized before the IMU service had completed its initialization.
The journal reported:
WARNING: could not read IMU pipe info JSON for 'imu_apps'
In those conditions, OpenVINS used:
prop_window: 0.075
When we ensured that the IMU service was fully ready before restarting OpenVINS, the system correctly identified the ICM42688 and used:
prop_window: 0.100
We reproduced this behavior multiple times.
However, correcting the startup sequence did not consistently eliminate the metric-scale inconsistency.
We would appreciate clarification on whether this startup behavior is known in OpenVINS 0.6.1 and whether more recent software releases improve service initialization or IMU detection.
We are not suggesting that the propagation-window difference has been proven responsible for the scale problem.
5. Timestamp-ordering anomalies
We also identified an independent timing anomaly during OpenVINS operation.
In one diagnostic session, OpenVINS reported an out-of-order VIO pose with a timestamp regression of approximately 33.3 milliseconds.
This corresponds approximately to one camera-frame period at 30 Hz.
The journal subsequently reported:
[HEALTH] ERROR: Sensor measurements with bad timestamps
Followed by:
AUTO-RESET RECOMMENDED Error code(s) detected 0x8000
OpenVINS then performed an automatic hard reset.
Our inspection of the corresponding raw camera and IMU timestamp streams did not reveal the same backward timestamp regression.
This suggests the possibility of an ordering problem within the subsequent VIO processing or publication chain, although we have not isolated its exact origin.
Importantly, we also observed incorrect metric-scale estimates during sessions without automatic resets.
Therefore, we consider the timestamp anomaly a separately documented problem, not a confirmed explanation for the scale inconsistency.
We would appreciate knowing whether this behavior is recognized in OpenVINS 0.6.1 and whether it has been addressed in later versions.
6. Visual-tracking quality
We also observed significant variation in the visual-tracking quality reported by OpenVINS.
Two diagnostic sessions conducted on October 1 produced the following measurements:
| Measurement | Session 14:16 | Session 14:24 |
|---|---|---|
| Estimated trajectory length | 11.81 m | 14.19 m |
| Final position closure error | 38.5 cm | 93.9 cm |
| Closure error relative to estimated trajectory | 3.3% | 6.6% |
| Median VIO quality | 81 | 30 |
These were independent hand-carried trajectories rather than identical controlled repetitions.
Nevertheless, they demonstrate substantial differences in reported tracking quality and final position consistency between sessions.
We also investigated OpenVINS camera-calibration settings, including temporarily disabling online camera-extrinsic calibration.
That experiment did not provide a consistent solution.
More recently, we restored the OpenVINS configuration using the official:
voxl-configure-open-vins starling2_2cam
We have not established whether the latest configuration reset changes the behavior.
7. Related ModalAI forum findings
Our investigation led us to several ModalAI discussions and more recent software developments.
We believe these findings may be relevant, although we have not demonstrated a causal relationship with our metric-scale problem.
A. OpenVINS initialization behavior on the same hardware
We reviewed the forum topic:
“voxl-mapper produces no map; VIO output rate halves to 15 Hz and does not recover until OpenVINS restart.”
This discussion involved the same Starling 2 Max C29 SKU and OpenVINS 0.6.1.
ModalAI engineering discussed an initialization-related mechanism involving static-jerk detection and camera-frame processing rates.
The discussion also clarified that apparently healthy VIO output can occur while the system is publishing blank VIO packets rather than actively estimating visual-inertial movement.
We found this particularly interesting because our estimated metric scale has varied substantially between diagnostic sessions.
We have not independently verified the specific 15 Hz behavior described in that thread.
However, it raises the following question:
Could differences in OpenVINS initialization or estimator activation contribute to inconsistent metric-scale estimation between sessions?
B. AR0144 synchronization and updated timestamp handling
We also reviewed the forum discussion titled:
“OpenVINS”
In this discussion, ModalAI referenced updated OpenVINS code and a synced-timestamp branch for synchronized AR0144 tracking cameras.
Our aircraft uses AR0144 tracking cameras, and we have observed timestamp-ordering anomalies.
We would appreciate clarification on whether those changes are applicable to our hardware and software configuration.
We are particularly interested in whether OpenVINS 0.6.1 has known limitations involving synchronized camera processing that have been addressed in later releases.
C. OpenVINS 0.7.8 and SDK 1.8
We also reviewed the SDK 1.8 release notes, including the changes introduced in:
voxl-open-vins-server 0.7.8
We understand that the newer implementation includes improvements involving:
- Asynchronous multicamera processing.
- Timestamp-ordered camera-frame handling.
- Visual-inertial initialization.
- Calibration and diagnostic functionality.
These changes appear relevant to two areas of our investigation:
Initialization consistency and timestamp ordering.
We are not assuming that OpenVINS 0.7.8 resolves our specific observations.
However, we would appreciate confirmation of whether these improvements address known limitations of the earlier implementation and whether the newer version is officially supported and validated for Starling 2 Max C29.
8. Request regarding the purchase of another Starling 2 Max
This is the main reason for contacting ModalAI.
We intend to purchase a second Starling 2 Max with the same hardware configuration as our current aircraft.
We will retain our existing drone.
Our priority is to receive the new aircraft with the most reliable and recent complete software configuration that ModalAI has officially validated for this specific model.
We do not necessarily want the highest available software version.
We want the most appropriate manufacturer-supported combination of SDK, OpenVINS, camera services, IMU services, PX4 integration, and supporting components.
We noticed that the published compatibility information identifies SDK 1.6.4 as a flight-tested release for Starling 2 Max D0012, while newer SDK releases contain significant OpenVINS developments.
We would appreciate direct clarification on the following questions.
1. Recommended software baseline
What is the newest complete VOXL SDK release currently recommended and officially flight-tested by ModalAI for Starling 2 Max C29?
Is there a newer validated configuration than the software installed on our current aircraft?
2. Factory delivery with validated software
Can ModalAI deliver a newly purchased Starling 2 Max with that recommended software already installed, configured, calibrated, and verified?
We would prefer a complete manufacturer-supported configuration rather than independently upgrading individual services.
3. Metric-scale consistency
Is ModalAI aware of any initialization, observability, calibration, or estimation behavior in OpenVINS 0.6.1 that could explain substantial variations in metric scale between sessions?
Have any related issues been corrected or improved in newer releases?
4. Timestamp and multicamera synchronization
Are the changes associated with the synced-timestamp branch or the newer asynchronous multicamera architecture relevant to our AR0144 configuration and the approximately 33.3 ms timestamp regressions we observed?
Is a corrected implementation included in a validated software release for this aircraft?
5. OpenVINS 0.7.8 compatibility
Is OpenVINS 0.7.8, or a later equivalent, officially supported and flight-tested on Starling 2 Max C29?
If not, what version does ModalAI recommend?
6. Factory calibration and verification
Can the new aircraft be delivered with its original camera intrinsics, camera–IMU extrinsics, IMU calibration, and confirmation of the exact software versions installed?
Does ModalAI perform VIO initialization and basic odometry verification before shipment?
7. Recommended configuration for our application
Given the hardware, software versions, and observed behavior described above, what complete software configuration would ModalAI engineering recommend for reliable autonomous indoor navigation?
We are particularly interested in avoiding known initialization and synchronization limitations.
9. Final objective
Our objective is to establish a reliable, repeatable, manufacturer-supported VIO baseline.
We have observed significant metric-scale inconsistency, along with separate initialization, timestamp-ordering, and tracking-quality anomalies.
We have also identified software developments that appear relevant, but we have not demonstrated that they resolve our particular problem.
We are not concluding that our existing aircraft has defective hardware or that a specific software bug has been proven responsible.
We would like ModalAI engineering to clarify which of these behaviors are known, whether newer validated software addresses them, and what configuration should be installed on a newly purchased Starling 2 Max.
We intend to keep our current aircraft and purchase a second identical unit.
Ideally, the new aircraft would arrive with the most recent officially validated software configuration, including any confirmed improvements relevant to OpenVINS initialization, camera synchronization, and metric-scale reliability.
We would appreciate a clear technical recommendation and confirmation of whether this factory configuration can be arranged as part of the purchase.
Thank you for your time and assistance.
Best regards,
Jorge Padilla