FASTNAV Far Obstacle Detector (FOD)
- The paper presents FASTNAV FOD as a lightweight, vision-based detector that extends obstacle detection range to 20 m, enabling continuous rover motion at 0.7–1.0 m/s.
- It employs a streamlined YOLOv8n model to process onboard NAVCAM imagery, integrating detections into a Far Traverseability Map for dynamic mode switching and proactive hazard mitigation.
- Field experiments in Mars-analogue settings demonstrate up to a 600% speed improvement and 900% range enhancement over conventional stereo systems, underpinning robust autonomous navigation.
FASTNAV Far Obstacle Detector (FOD) is a lightweight, onboard, long-range vision-based obstacle detector for planetary rover navigation. Introduced as the core perception module in the FASTNAV autonomy system, it is designed to extend hazard perception from the short ranges typical of stereo-based systems to about 20 m, thereby enabling continuous rover motion at roughly 0.7–1.0 m/s while preserving safety through confidence filtering, fallback behaviors, and integration with a Far Traverseability Map used by the guidance, navigation, and control system (Luna et al., 7 Oct 2025).
1. Operational role and navigation concept
FASTNAV FOD addresses a central limitation of current planetary rovers: they typically move at only a few centimeters per second because they must frequently stop, observe nearby terrain, and replan. The broader FASTNAV framework is explicitly organized around a multi-range, multi-mode navigation concept in which the rover can switch from a conservative low-speed mode to a higher-speed “FASTER” mode when long-range hazard perception is available. FOD provides that long-range perception by detecting obstacles far enough ahead that the rover can continue moving while still having time to replan or slow down (Luna et al., 7 Oct 2025).
The system is framed as a departure from the traditional Sense-Model-Plan-Act pattern, whose stop-and-check behavior constrains mission productivity. Conventional stereo systems are described as often providing a horizon of only about 2–3 m, whereas FOD is intended to detect hazards at around 20 m. The paper relates detection distance to reaction time through
With m and m/s, the rover has about 28.6 s of reaction time, which is characterized as roughly an order-of-magnitude improvement in planning horizon relative to 2–3 m detection ranges (Luna et al., 7 Oct 2025).
This extended look-ahead is the basis for sustained high-speed driving. The stated operational consequence is not merely earlier recognition of a rock, but the ability to adjust heading, trigger a controlled slowdown, or switch modes before the rover reaches the hazard. A plausible implication is that FASTNAV treats perception range as a control-enabling quantity rather than as a passive sensing metric.
2. Perception architecture and data flow
Architecturally, FOD is a machine-learning-based visual obstacle detector built around a lightweight YOLOv8n object detector. The choice of YOLOv8n is motivated by small model size and low inference latency, making it suitable for low-power onboard hardware. The input is raw NAVCAM imagery acquired on board the rover, and the detector produces bounding boxes with confidence scores (Luna et al., 7 Oct 2025).
These detections are then transformed into rover-relative geometry and encoded into a Far Traverseability Map used downstream by the guidance, navigation, and control system. In this sense, FOD is not just a standalone classifier: it is a perception layer that feeds navigation logic. Operationally, detections are filtered for confidence, geometrically transformed, and published at a configurable rate of 1 Hz by default, up to 5 Hz (Luna et al., 7 Oct 2025).
The architecture includes explicit mitigations for both principal failure modes. False negatives are handled by falling back to the conservative RAPID mode, while false positives are reduced through confidence thresholding. The detector therefore participates in a safety-aware autonomy stack rather than in an isolated perception benchmark. Its stated role is to support early replanning, dynamic speed scaling, and safe mode switching (Luna et al., 7 Oct 2025).
3. Training corpus and robustness strategy
FOD was trained and fine-tuned on a composite rock-obstacle dataset centered on a single-class “rock” detection problem. The base corpus was ESA’s Katwijk dataset, from which 3000 images were manually labeled. This was augmented with additional field imagery from Upwood Quarry (300 images) and the Bardenas Desert (300 images) (Luna et al., 7 Oct 2025).
Standard YOLO augmentations were applied, specifically scale, rotation, brightness changes, and synthetic dust. These augmentations were introduced to improve robustness to harsh illumination, flares, and scene anomalies. In the planetary navigation setting described by the paper, this is directly tied to the variability of surface imagery under changing sun angles and dusty conditions (Luna et al., 7 Oct 2025).
The training design indicates that robustness was pursued through data composition rather than through a specialized detection taxonomy. The detector is narrowly targeted at rock-like hazards in rover imagery, but the augmentation policy suggests an attempt to stabilize performance under operationally adverse visual conditions rather than optimize only nominal-image accuracy.
4. Integration with FASTNAV autonomy
Within FASTNAV, FOD functions as the perception enabler for high-speed traversal without sacrificing safety. The paper positions it inside a multi-mode behavior stack in which long-range detections support proactive navigation decisions. This includes early replanning, dynamic speed scaling, and safe transitions between conservative and higher-speed modes (Luna et al., 7 Oct 2025).
The reported system-level characteristics tie FOD to the rest of the autonomy pipeline rather than treating it as a detector in isolation. The system reported 20 m average obstacle detection range with 95% reliability, 2 Hz continuous path planning updates, trajectory-following error below 0.3 m RMS, and emergency stop distance below 1.5 m at maximum speed (Luna et al., 7 Oct 2025).
The paper further describes FASTNAV as part of a broader suite of AI-enabled systems while emphasizing that, within rover navigation, FOD supports a progressive enhancement path: it can be integrated with traditional rover control architectures without a full redesign. This suggests a deployment strategy in which long-range perception is added as an enabling layer over existing navigation logic, thereby lowering adoption risk for mission architectures that already rely on conservative control stacks (Luna et al., 7 Oct 2025).
5. Field validation and quantitative performance
Experimental validation took place in a Mars-analogue environment at Bardenas Reales, Spain. The results are used to support the claim that FASTNAV/FOD is ready at Technology Readiness Level 4. The field trials report speed and uptime for several operating modes (Luna et al., 7 Oct 2025).
| Operating mode | Uptime | Average speed |
|---|---|---|
| 0.7 m/s autonomous | 60.95% | 0.43 m/s |
| 1.0 m/s autonomous | 35.94% | 0.36 m/s |
| 1.2 m/s teleoperated | 95+% | 1.14 m/s |
| Baseline conventional | 90+% | 0.10 m/s |
The degradation at the 1.0 m/s autonomous setting is attributed mainly to frequent point-turns, especially when sharp turns with radius below 2 m triggered point-turn behavior, lowering average traverse speed. Under the tested conditions, the paper identifies 0.7 m/s as the optimal balance for autonomous operation, providing about a sevenfold improvement over conventional operations while maintaining safety margins (Luna et al., 7 Oct 2025).
The quantitative comparisons are also expressed as mission-level deltas. Comparing 0.7 m/s autonomous operation to a 0.1 m/s baseline, the paper gives
Comparing 20 m detection to 2 m detection, it gives
These are associated with traverse projections of 300–600 m per sol versus 50–100 m per sol for conventional operations, and about 40% more scientific investigation time due to faster transit between targets (Luna et al., 7 Oct 2025).
A common misunderstanding is to equate the headline capability of travel at or above 1.0 m/s with the achieved autonomous average speed in field conditions. The trial data show a more conditional result: FOD enabled sustained high-speed operation in principle, but achieved autonomous performance depended on path geometry, mode-switching behavior, and the frequency of point-turn events.
6. Relation to broader far-obstacle detection research
FASTNAV FOD belongs to a broader class of far-obstacle detectors, but its formulation is specific: it is a camera-based planetary rover module that converts raw NAVCAM imagery into obstacle detections and then into a traverseability representation for GNC. Other recent systems illustrate alternative sensor and inference strategies for the same general problem of extending the planning horizon.
In radar-only autonomous driving, NVRadarNet constructs a top-down bird’s-eye-view input from temporally accumulated detections from 8 radars covering 360°, operates over an 800 × 800 grid spanning ±100 m at 25 cm/pixel, predicts obstacle classes, oriented geometry, and free-space occupancy, and runs faster than real time on an NVIDIA DRIVE AGX embedded GPU with 1.5 ms end-to-end inference (Popov et al., 2022). This shows that far-obstacle detection can be realized without cameras, and that occupancy/free-space outputs can expose unclassified hazards.
In off-road robotics, radar-only local navigation has been demonstrated by adapting G-VOM to radar through thresholding, voxelization, intensity-weighted ground estimation, and obstacle classification from solid voxels above the ground plane. That system reports radar returns beyond 250 m, a real barrier detected in radar at 80 m and not visible to lidar until within 40 m, and successful autonomous navigation at 2.5 m/s over 350 m using only radar for ground plane estimation and obstacle detection (Overbye et al., 2023). This suggests that the central FOD design question is often not detection alone but how long-range evidence is converted into a navigation cost structure.
A different modality appears in small AUV navigation, where a low-cost forward-looking sonar with a nine-beam prototype, 219 distance bins per beam, and range out to 50 m ahead is used to build a probabilistic local polar map. There, collision avoidance is formulated through posterior expected loss that explicitly couples detection, avoidance, and waypoint planning, and field trials on the 690 AUV are reported to have successfully avoided collision in all five runs (Morency et al., 2023). Relative to FASTNAV, this emphasizes a decision-theoretic treatment of uncertainty rather than a box-detection pipeline.
Monocular vision-based far-obstacle detection also has a distinct lineage. Fully convolutional monocular depth estimation has been used for long-range obstacle perception with dense depth prediction, synthetic training up to 200 m, and reported inference around 300 Hz in the paper’s reporting context, explicitly targeting fast, robust, long-range obstacle perception rather than precise 3D reconstruction (Mancini et al., 2016). FASTNAV FOD differs in producing rock detections and rover-relative geometry rather than a dense depth field, but both approaches are motivated by the same requirement: extending reaction distance for high-speed autonomous motion.
The acronym “FOD” can also denote foreign object debris in runway surveillance rather than Far Obstacle Detector in rover autonomy. In runway FOD detection, domain-randomized synthetic augmentation with YOLOv5m, BoxInst-generated masks, 2227 synthetic images derived from 461 extracted masks and 29 runway backgrounds, and an out-of-distribution evaluation increased mAP@0.5 on the OOD test set from 41.8% to 92.6% (Farooq et al., 2023). The overlap is terminological rather than architectural. FASTNAV FOD is a rover navigation perception layer; runway FOD detection is a tiny-object detection problem under domain shift.
Taken together, these adjacent lines of work suggest that “far-obstacle detection” is not a single algorithmic family but a systems problem spanning sensing modality, map representation, risk handling, and planner integration. FASTNAV FOD’s specific contribution is to show how a lightweight YOLOv8n detector on raw rover NAVCAM imagery can be tied to mode switching, traverseability mapping, and safety-aware high-speed planetary driving (Luna et al., 7 Oct 2025).
