Under canopy, where GNSS is unreliable, published work fuses a range sensor (lidar or 4D imaging radar) with an IMU, and orchard robots also use cameras. Lidar loses penetration in very dense canopy, radar blurs thin trunks and loses signal at long range, and cameras lose features to occlusion, poor exposure and motion blur. So the suite should be chosen through a trade study against the vehicle’s weight, power and cost budget. It should then be proven by replaying recorded outage logs and trialling it under canopy as dense as the real site. A datasheet cannot do that job.
The same goal raises narrower questions: how early a link dropout can be predicted, where sensors go on the chassis, whether a subsea housing will survive, and when the vehicle should stop and wait for a fix. Each needs its own evidence.
Which sensors hold position when the canopy blocks GPS?
Published canopy and row-crop work uses lidar or stereo cameras for geometry, with an IMU or odometry for motion. Fusion back ends combine them, and some add radar, as in radar-lidar-inertial and baro-radar-inertial odometry. Which units, in which mounting layout, fit a given airframe’s weight, power and cost budget is something you measure for that vehicle and that canopy.
TechnoLynx starts this work from the vehicle’s budget. TechnoLynx records the vehicle’s weight, power and cost limits alongside the canopy density and the localisation accuracy the mission needs. TechnoLynx then shortlists lidar, stereo and IMU combinations and a VIO or SLAM fusion design, integrates and tunes the pipeline on the target compute, and tests it in representative canopy trials against that budget. This work sits within TechnoLynx’s AI engineering for robots, drones and autonomous machines.
How each modality behaves under canopy in published field tests
| Modality | Measured result under canopy or in farmland | Where it breaks down | Role in the suite |
|---|---|---|---|
| 3D lidar | A ground robot followed crop rows over the canopy without GPS at 3.55 cm average deviation from the row centre line (Velodyne VLP-16, arable fields). Lidar-inertial methods gave the most stable, consistently accurate results on the AgriLiRa4D farmland UAV benchmark. | Beams are intercepted by the lower canopy in very dense crops such as soybean. The Viking Hill lidar was out of range at 30 m to 50 m in forest, because forest surfaces reflect poorly. Rain, snow and fog scatter or absorb the beam. | Geometry sensor for over-canopy row following |
| 4D imaging radar | Segmentation IoU close to lidar for ground (91%) and canopy (86%) in the Viking Hill forest dataset. Baro-radar-inertial odometry held 0.5% to 3.2% drift per distance travelled in forest and city flights. | Thin structures: trunk IoU of 56% against 74% for lidar. Rain and snow attenuate the signal. Radar-inertial odometry failed on nearly all flat-farmland sequences. | Backup when optics are blinded. The Viking Hill authors expect mud, dry snow and dust on the radome only partly to reduce signal-to-noise ratio; they have not yet tested this |
| Stereo or visual camera | Template-based orchard localisation kept lateral error under 5% of row spacing (under 2% with visual odometry) across winter, spring and summer foliage. | Changing light, repetitive crops with little texture, motion blur, vegetation moved by rotor downwash, occlusion and badly exposed frames. | Geometry and visual odometry source in orchard localisation |
| IMU | Needs no external signal, so it is useful where GPS is denied. | Drift from integrating acceleration. On the NCLT ground-robot data, lateral error grew quadratically with blackout duration. | Fused with GPS to limit drift |
The budget sets the direction of the study. If the vehicle can carry lidar, stereo and IMU together, the trade study picks the specific units and layout for canopy-like conditions. If it cannot, the study turns to a stereo or visual-inertial suite. That means accepting less robustness in the densest canopy, and the reduced suite still has to pass the same tests before the build is committed.
What added redundancy costs in grams and watts depends on the units, the mounts and the compute that runs the fusion. Compute belongs in that cost too. RadarOcc, which processes raw 4D radar tensors, ran at about 3.3 frames per second on a single desktop GPU and about 7.46 after optimisation, and its authors point to embedded GPUs such as Jetson Orin for future real-time use. Every added sensor also adds mass and power draw, which can shorten flight time or reduce payload even when localisation gets better. How the fused estimate is structured, and which edge platform runs it in real time, is a separate design problem from choosing the sensors.
Has anyone proven this suite in conditions like ours?
Partly. The published evidence cited here comes from single sites or datasets: apricot rows in Davis, a set of arable fields, one forest site recorded twice, and farmland near Nanjing. Earlier results lower the risk but do not transfer automatically. A suite proven at one canopy density may not carry over to a denser one. A lidar study saw this pattern: its methods worked in corn but were unreliable in dense soybean. That study measured plant traits, not navigation. Plan the ground truth data as part of the test budget too. Under forest canopy, even post-processed RTK reached sub-centimetre accuracy at only a few spots.
Will adding lidar or radar fix drift between rows, or is it a filtering problem?
Not always: some lock loss between rows is a modelling problem, so check the filter before buying hardware. Predicting a link dropout rests on monitoring link-quality signals. Choosing what hardware to add calls for a trade study. Proving graceful degradation calls for a replay test, and its measured drift should set the stop-or-continue threshold.
TechnoLynx starts by framing which of the three questions a team has, against the robot’s operating environment. TechnoLynx then builds the matching tool on the target embedded compute: anomaly or fault detection, a sensor-suite trade study, or a scenario-replay test harness. TechnoLynx validates it against logged failure and outage data and sets the alert lead time and the stop-or-continue drift threshold from those measurements.
Which early signals predict a sensor-link dropout in time to recall the drone?
Radio link failure can be predicted from link-quality measurements before the link goes. Boutiba, Bagaa and Ksentini (EURECOM) recorded RSRQ, RSRP, PHR and CQI on an OpenAirInterface 5G testbed. They found that RSRQ, PHR and CQI correlated with connectivity, and they reported 98% accuracy in predicting connection status. Their framework sends the prediction to an application that modifies the UAV’s trajectory to avoid link failure. For telemetry radios, one radio guide gives fade margin as the live range indicator. Subtract noise from RSSI and halve the result to get a margin in decibels; each 6 dB of margin doubles range.
Two limits apply. A 5G testbed is not a drone mesh, and lead time rests on how well the logged pre-failure signals represent real failures. A failure mode the model has never seen can still arrive without warning. Battery sag is a plausible input but an unproven predictor of link loss, so test it against the logged data before writing it into an alert rule.
Does a better IMU buy enough time to get through a row?
Usually not on its own. On the NCLT dataset, the FusionCore unscented Kalman filter drifted 1 to 2.6 m/km when GPS blackouts stayed under 203 s. When blackouts lasted 240 to 462 s, drift rose to 6 to 10 m/km, and lateral error grew quadratically with blackout length. A baseline filter that rejected bad GPS fixes and fell back to wheel-encoder dead reckoning drifted 31.84 and 50.11 m/km on two sequences. Baro-radar-inertial odometry held 0.5% to 3.2% drift per distance travelled in forest and city flights.
Worked example, assuming drift grows in proportion to distance (a simplification). A 1.8 m-wide robot straddling rows 0.75 m apart has 22.5 cm of lateral tolerance before it damages the crop. At 0.5% drift the error reaches 22.5 cm after 45 m (0.225 m ÷ 0.005). At 3.2% it gets there after about 7 m (0.225 m ÷ 0.032). Lidar crop-row following held an average deviation of 3.55 cm from the row centre line. That comparison has to come from replaying your own outage logs.
FusionCore estimates gyroscope, accelerometer and wheel-encoder yaw-rate biases online as part of a 23-element state. With the same sensor data, it cut absolute trajectory error on ten of twelve NCLT sequences compared with its baseline, by 1.2× to 22.2×.
How do we prove the suite degrades gracefully, and when should it stop and wait?
Replay recorded canopy outage logs through the fusion stack with realistic sensor timing and faults injected. Measure how drift and the filter’s own reported uncertainty change, then set the stop threshold from the measured drift. Timing realism matters. One sensor-emulation framework requires emulated latency to follow the real sensor’s latency distribution. It added 73 ms to its 27.42 ms nominal end-to-end latency to match the 100 ms set for a typical uBlox receiver in PX4. PX4’s EKF compensates for sensor delay only when the nominal delay is specified in its configuration. Missing fixes are not the only fault to inject. One NCLT sequence contains 105 fixes, each 720 to 840 m from ground truth, that arrive as valid quality-3 GPS fixes. A deeper treatment of harness design and rare edge cases is in proving fusion and alerting systems through logged-data replay.
A replay check for graceful degradation
- Assemble outage logs with ground truth wherever it exists, knowing that under canopy RTK gives precise fixes at only a few points.
- Replay them through each candidate configuration, with latency and jitter drawn from each real sensor’s measured distribution.
- Inject blackouts of increasing length, plus fixes that look valid but are wrong.
- Record drift rate against blackout length. Check that the filter’s reported covariance grows with the true error, because silent drift means error grows while confidence does not.
- Divide the mission’s lateral tolerance by the measured drift rate to get a distance budget, and turn that budget into failsafe settings.
PX4’s position-loss failsafe triggers when the estimate becomes invalid. That happens in two ways: a timeout since the last fused sensor that measures speed or horizontal position directly (GNSS, optical flow, airspeed, VIO, auxiliary global position), or, on hovering vehicles, estimated horizontal inaccuracy above COM_POS_LOW_EPH. The PX4 guide also lists COM_POS_FS_EPH as the horizontal error threshold for multicopters and VTOLs in hover (infinite on fixed-wing), plus a maximum inertial dead-reckoning time. In autonomous modes a multicopter then falls back from Return (needs global position) to Land (needs local position) to Descend. The drift rate measured in replay is what should set them.
How do we justify radar-plus-camera over lidar on cost and coverage?
You need per-unit cost, range and field of view for each candidate set, plus at least one measurement in canopy or rows rather than open-field driving.
Numbers for the investor slide
| Figure | Lidar | mmWave radar | Source and caveat |
|---|---|---|---|
| Unit cost (industrial) | About $300 to $5,000 | About $200 to $3,000 | Industrial sensor supplier QJKH. Depends on quantity, model and quote |
| Range (industrial) | 0.1 to 200 m | 0.2 to 300 m | Same supplier’s specifications |
| Range (automotive) | Within 150 m | Within 250 m | Review of mmWave radar and vision fusion |
| Field of view | 270° to 360° | 60° to 120° | Supplier specifications |
| Accuracy | ±2 to 10 mm | ±50 to 200 mm | Supplier specifications |
| Velocity | 3D point cloud only, no velocity output | Doppler, 0.1 m/s resolution | Supplier specifications, automotive review and 4D radar survey |
The two range figures come from different markets, industrial and automotive. Coverage changes the cost picture. For an indoor pick zone, the same supplier argues that one 270° scanner does the work of three 90° radar nodes, and that installed cost includes cabling, IP-rated connectors, controller inputs and programming hours as well as unit price. The case for radar plus camera rests on complementary strengths. Radar is strongest at distance, radial speed and all-weather operation, while the camera adds classification and angular resolution. Most automotive manufacturers use radar with cameras, but that is open-road evidence. For a technically literate investor, the forest and farmland rows in the modality table above should carry more weight than any automotive figure. So should current quotes, since prices depend on quantity, model and supplier.
Comparing configurations in one field trial before the next build
You can, if every candidate sensor rides on the same rig, is logged in sync, and the configurations are compared offline. AgriLiRa4D logged 3D lidar, 4D radar and IMU together, synchronised over IEEE 1588 PTP. It refined lidar-radar calibration with corner reflectors and then benchmarked lidar-inertial, radar-inertial and radar-lidar-inertial pipelines on the same flights. The ranking depended on terrain. Lidar-inertial methods were the most consistent, while the radar-lidar-inertial method did best on terraced farmland and often failed on hilly ground. Attribution still needs care. Higher altitude cut point density and the overlap between lidar and radar, weakening the cross-modal constraints. A shared clock is a dependency of its own, and watching it is a separate monitoring job.
Will the housings and mounts survive, and how do we test across seasons?
For subsea vehicles, pressure-test every housing, monitor humidity inside it, and avoid fastenings that starve stainless steel of oxygen. Treat any claim about unattended service life as unproven until it has been tested. For canopy vehicles, replay logs from each season against each candidate configuration before buying a fleet.
Sealing, corrosion and fouling on subsea housings
On the MARVEL autonomous surface vehicle, built for GNSS-denied fusion testing, each electronics housing was submerged and pressure-tested. Humidity, temperature and pressure inside the enclosures were monitored, and any weak point was fixed before in-water trials. An open underwater robotics toolkit tested a sealed joint up to 5 bar, roughly 40 m of seawater, including lateral loads on a shaft sealed with two O-rings. A humidity sensor inside gives early leak warning. On seabed tripods at 55 to 75 m in Admiralty Inlet, with near-bed currents approaching 3 m/s, about half of the stainless fasteners corroded close to mechanical failure in one deployment. The common cause was nylock nuts, and switching to stainless nuts with lock washers stopped the widespread corrosion. Zinc anodes reduced but did not eliminate oxidation of steel, and a high-copper antifouling coating kept fouling minimal.
The coupon deployments cited here lasted only three to ten months. One three-year wave-energy failure log found that corrosion, fatigue and wear were over-predicted for the early deployments, because those mechanisms act late in service life. A short trial says little about month eighteen.
Testing across canopy densities and seasons before a fleet purchase
Results can differ between seasons. The orchard template method was tested in dormant winter, sparse spring and dense summer foliage. In the Viking Hill forest data, mean per-tree IoU dropped in every trunk-size class from the May session to the June session with full undergrowth. For thin trees, mean per-tree IoU fell from 65.1% to 57.8%. Replaying existing drift logs against each configuration is the cheap first filter, and only configurations that pass should go to limited field trials. Replay cannot reproduce canopy or seasonal conditions missing from the logs, so some field validation always remains.
Where should lidar or radar go so branches don’t knock it off?
The published evidence settles field of view and vibration far better than it settles branch strikes. The cited work documents two placement patterns. For navigation over the canopy, a lidar mounted centre-front at 1.5 m and tilted down collected enough crop-top points to detect rows. For navigation under the canopy, a forward-looking 3D sensor placed near the ground, below the treetops, sees the tunnel of ground and canopy roughly 10 to 20 m ahead. The ZED stereo camera in that work sat front-centre, facing forward.
Vibration is measured. A lidar mounted above a tractor’s hood, far from the centre of gravity, picked up both small high-frequency and large low-frequency vibration. Its mean positioning error rose from 20 cm at 2 km/h to 35 cm at 7 km/h. A stabilising system of twelve silicone-gel isolators cut mean error by 57%, from 38 to 17 cm, and support bars cut it by 41%, from 34 to 20 cm. Any IMU sharing that mount picks up the same motion, so sensor placement and IMU isolation have to be decided together, with the airframe or chassis in view.
For branch strikes, keep the sensor inside the vehicle’s swept envelope and use a guard that stays out of the beam. A guard changes occlusion, so it gets tested in replay like any other change.
On IMU redundancy: extra IMUs allow a failed sensor to be detected and isolated, and they let the error covariance be estimated from the redundant measurements. Redundancy also raises the problem of sensor placement and orientation. A 192-element inertial array fell short of its theoretical accuracy, and the likely causes were scale-factor errors, sampling misalignment and uncertainty in sensor position. If the mission must detect and isolate a failed IMU in flight, fit redundant units. If not, one well-isolated, well-calibrated unit is the simpler choice.
Who can review a subsea mounting plan before fabrication?
Use a classification society where formal certification or insurability is required, and a specialist marine consultancy where you need an engineering review. In both cases, use the standard for the asset class as the checklist, plus published post-mortems for failure modes it misses. DNV-ST-0164 for tidal turbines (Edition 2015-10, amended 2021-11) sets minimum requirements covering structures, safety and control systems, electrical systems and corrosion protection. It also serves as a contractual reference and as the basis for DNV verification. The DNVGL-SE-0163 certification specification, published in October 2015, lays out the path from prototype to type and project certification, and it names investors and insurers among its readers. Both are tidal turbine documents. Whether a society will review a vehicle’s sensor mount as a standalone job is a question to put to the society directly.
Post-mortems add field failure data. Albatern’s three-year failure log from wave deployments on the west coast of Scotland found human error behind 27% of recorded failures, against 17.7% in its FMEA. Major structural and hydraulic failures came early, and minor electrical and instrumentation failures came later. A European Commission study of ocean energy development reviewed failures and good practice in wave and tidal technology and called for an integrated approach that learns from past experience. A tidal measurement best-practice manual recommends two instruments rather than one in case one is damaged, citing IEC guidance from 2013. It also warns that uncompensated metal structures near a compass can make heading data useless. Check the reviewer’s direct subsea deployment experience before you commit.
FAQ
What can block a GNSS signal?
Tall tree foliage can block GNSS signals or cause multipath errors, which makes satellite positioning on a moving receiver unreliable or impossible inside orchard rows. Indoor spaces and urban canyons degrade it in the same way.
What is sensor fusion in autonomous vehicles?
Sensor fusion combines sources such as GNSS, IMUs and velocity logs through algorithms that estimate the vehicle’s state in real time, especially when GNSS is unavailable. It also keeps position and orientation estimates going when one sensor fails.
