When NASA engineers want to test new ideas for Mars exploration, they must first create conditions as faithful as possible to "Martian" ones here on Earth. That is why for decades they have been returning to the deserts of the American West – especially California's Death Valley and the wider Mojave – where bare dunes and rugged volcanic slopes offer an ideal, yet unforgiving proving ground. This year, that landscape once again became an open-air classroom: a team from the Jet Propulsion Laboratory (JPL) conducted two intensive field campaigns, in late April and early September 2025, with three research drones equipped with a new generation of autonomous navigation software. It involves a suite of solutions being developed at JPL under the program name Extended Robust Aerial Autonomy – an ambitious effort to ensure future aircraft above Mars fly reliably even where monotone dunes "deceive" them, and land safely even where the terrain is packed with rocks.

From Ingenuity to "more robust" autonomy

The impetus for the new wave of testing came directly from the lessons left by Ingenuity, the small helicopter that proved from 2021–2024 that powered flight is possible in the thin Martian atmosphere. In the final stages of its mission, during a series of flights over uniform, gently undulating dunes, Ingenuity's visual odometry was occasionally misled: the ground "looked" similar everywhere, contrast was poor, and the algorithm had difficulty estimating actual speed and displacement. On January 18, 2024, on its 72nd flight, the helicopter experienced a rough contact with the ground and damaged its rotor blades; on January 25, NASA confirmed the end of its flight phase. Although the aircraft spectacularly surpassed the plan (72 flights instead of the few planned), it was precisely the monotone dunes that – paradoxically – taught the most important lesson: autonomous systems must "read" even the dullest possible scene.

Why Death Valley and Mojave?

Located in the heart of the North American deserts, Death Valley has been NASA's "proxy Mars" for decades. Ever since the 1970s, when field preparatory measurements for the Viking missions were made there, this area has offered two extremely different worlds, equally difficult for navigation: the bare, uniform dunes of Mesquite Flat Sand Dunes and the rocky slopes of an area researchers colloquially call Mars Hill. On the dunes, software easily loses landmarks because sand patterns are repetitive and poor in detail. On Mars Hill, however, there are too many "obstacles": edges, shadows, and rough stone slabs test the system's ability to recognize safe landing zones in flight. Precisely this "empty – overcrowded" combination makes Death Valley an ideal proving ground for teaching systems that must work both where the camera sees too little and where it sees "too much".

Flight campaigns 2025: two terms, one mission

During two campaigns – in late April and early September 2025 – the JPL team, with special permits from the park administration, performed a series of short flights in morning and late-afternoon "windows" when thermals and wind are more predictable. Temperatures reached 45°C, so the schedule, logistics, and equipment protection were part of the engineering problem. Under temporary shelter (a ventilated tent), telemetry and video streams were monitored, and after each flight followed a quick analysis and a relaunch with modified parameters. For additional terrain variation, part of the tests was moved to Dumont Dunes in the Mojave – a place where NASA checked Curiosity rover mobility back in 2012 – to expose the software to rhythmic, irregularly spaced sand waves that easily "trick" the algorithm.

Three drones, three roles

To speed up iterations, three differently configured platforms were used. The "sensor mule" carried multiple cameras and a set of interchangeable optical and polarization filters; the goal was to examine how individual spectral windows enhance local contrast on sand and help detect microstructures. The second aircraft was the "computing racer" – with faster processing at the edge (edge computing) where variants of visual odometry, feature detection algorithms, and risk assessment ran. The third drone was the "baseline" – a reference for comparisons – with settings that were minimally changed so that every change could be attributed precisely to what was actually being tested.

What exactly does "extended robust aerial autonomy" mean?

At the center of the approach is multi-level sensor fusion and "awareness of one's own uncertainty". At the lowest level, visual odometry combines camera frames with inertial measurement unit (IMU) data to estimate speed and position. But as soon as the algorithm recognizes that the scene is losing informativeness – for example, because series of sand ridges repeat and shadows are short – the system increases reliance on alternative signals (barometer, wind models, flight dynamics constraints). Additionally, it can perform a short "pop-up" maneuver: climbing a few meters to briefly "change perspective", capture relief with higher contrast, and reset accumulated error. At a higher level, landing risk assessment works: semantic segmentation of the scene (sand, rock, shadow, tracks) and a quick measure of "roughness" generate a map of candidate zones, and planning selects the one that satisfies safety and scientific criteria.

Geologists as software teammates

Field geology here is not decoration but part of the algorithm. Geologists mapped sand types, prevailing wind directions, and micro-relief "traps" so that telemetry could be linked to the processes shaping the dunes. If a drone needs to search for traces of sediment that retains water longer or fine dust with potentially interesting mineral signals, the system must know when it is worth investing energy in flying over "difficult" zones and where a safe landing is most likely. In practice, this is a synergy of science and navigation: the risk map and the geological interest map are created together and compete for the same goal – better science without unnecessary risk.

What practice brought: concrete gains

Tangible progress was recorded after just two campaigns. Filter combinations were established that improve ground tracking over uniform scenes; tactics of short "pop-up" ascents to reset error were validated; and new algorithms for selecting landing sites in "cluttered" scenes like Mars Hill showed greater robustness to shadows and geometric illusions. Procedures for "return from the edge" were also tested – what to do when the system detects a growth of uncertainty in its own position estimate – to prevent a domino effect of errors. Special attention was paid to short-lived sand vortices ("blowback") that can cover sensors: fast data cleaning and anomalous noise detection routines were developed.

Broader context: 25 technologies for the Red Planet

Flight tests are part of a broader portfolio of the Mars Exploration Program, which supported twenty or more development directions during 2024 and 2025 – from autonomy and communications to precision landing (EDL) and better "computing on the edge". The idea is clear: future robots must make more good decisions on the terrain itself, without waiting for instructions from Earth (which, due to signal delay, can take over 20 minutes one way), and the collected data should be of higher quality already "in the first pass". Plans also mention more advanced aerial platforms (e.g., concepts like the Mars Science Helicopter) and swarm systems with multiple smaller helicopters in different roles – from reconnaissance and mapping to communication relays and logistical delivery of small payloads.

Ingenuity as a model – and a warning

Ingenuity's legacy is twofold: it inspired a wave of new ideas, but it also very concretely showed where the limits of visual navigation lie. Analyses published in late 2024 confirmed that uniform dunes led to incorrect estimates of horizontal speeds at contact, which most likely resulted in damage to the blades. These insights have now been translated into a requirement for "more robust" autonomy: the system must know when its scene is "poor" and how to compensate for it, and landing procedures must be more tolerant of short-term errors.

Robot dogs in White Sands: what four-legged scouts do

The California desert was not the only place this summer. In August 2025, scientists and engineers from NASA's Johnson Space Center with partner universities spent five days on the gypsum dunes of White Sands National Park in New Mexico. There, they trained four-legged robots – "robot dogs" – for movement on loose, bright ground, fusion of LIDAR, stereo-vision, and inertia, and for basic scientific tasks such as recognizing layers and taking samples. Such platforms can be the first to enter more challenging terrains, map wind shelters, mark safe zones for aircraft landing, and set up temporary meteorological and communication nodes. In combination with drones, this is a symbiosis of ground and air that opens more ambitious exploration profiles to future missions.

Atmosphere that "crackles": why meteorology matters for flight too

Recent observations of electrical discharges ("mini-lightning") in the atmosphere of Mars, associated with dust devils, remind us that the environment is not a static backdrop. For aerial platforms, this means another input into the risk model: recognizing patterns that precede such phenomena, changing altitude, shortening the route, or postponing landing. Simultaneously, more robust computing (HPSC) is increasingly coming to aircraft, enabling more complex models in real-time, including supervised learning on the mission itself: the drone builds a "situation log" and through weeks of operation becomes better at predicting its own weaknesses.

Operational choreography: briefing – flight – analysis – iteration

In the field, everything looked like a small space mission. The day began with a briefing with a forecast of wind and insolation, definition of experiments, and division of roles. Short flights with clearly set goals followed, then immediate downloading of logs, synchronization of video frames and graphs, and quick statistical calculations: how much drift increased, what was the density of "reliably detected" features, where semantic segmentation was right, and where it erred. Successful scenarios returned to the air with minor corrections; problematic ones were reproduced in the simulator before the next change. Such a rhythm closes the learning loop and saves time in the field, which is precious in desert conditions.

Technique under the hood: sensors, filters, semantics

The greatest gains were shown where good hardware and smart software meet. Different optical and polarization filters helped highlight edges and microtextures of sand even when shadows are minimal. Semantic segmentation – dividing the scene into categories like "sand", "rock", "shadow", "track" – allowed the algorithm to ignore deceptive signals (e.g., dark shadows that look like "holes") and more safely estimate roughness and slope. Additionally, fast routines for "querying" its own uncertainty were introduced: if the model estimates that the error is growing above a threshold, the system seeks an additional informative frame or corrects the flight profile.

Logistics and environmental protection: how to fly in a protected area

Death Valley is a strictly protected national park, so flights are allowed only with special permits and strict protocols. This year, the JPL team received only the third such permit ever. Flights were temporally and spatially limited, corridors carefully defined, all to reduce the impact on visitors and the sensitive environment. At the same time, precisely this cooperation with the park administration emphasizes the importance of the place: Death Valley is not just a spectacle, but a living laboratory that helps understand desert processes on Earth – and worlds beyond it.

Looking ahead: from prototype to missions

What follows? In the short term, new algorithms are being "calibrated" on an increasingly rich set of flight data and compared with Martian observations (e.g., footage from Perseverance and orbiters). In the medium term, they enter demonstrators and concepts mentioned in the Mars Exploration Program plans for the next decade: advanced science helicopters, air-ground teams, and small payload logistics missions. In the long term, the goal is clear: robotic aircraft that fly further over Mars, land more safely, and perform more meaningful scientific tasks – even in the most ungrateful terrains. If autonomy manages to "read" sand without patterns and deftly choose landing sites among rocks, science will gain both breadth and depth that we could not plan until now.