Mission: Vesta

MISSION: VESTA

We’re building tools to plan science walks on Mars.

Begin the mission

Begin the mission

Two tools, one map of Mars. Start with the guided mission.

HiRISE map of lower Mount Sharp: habitat H, stops 1 to 3, the route from orbit in red dashes through rocks the scout found, the scouted route in white.

Vesta Scout · the guided mission

Look before you walk.

Pick science stops, plan the route, send a scout helicopter to find the rocks orbit can’t see, then walk.

Eye level on lower Mount Sharp: a red guide line runs over the ground to stop 1, labelled Next: Altadena drill site, 345 m.

Vesta Vision Early version

Read the ground.

The site in 3D, down to eye level: follow the red line to the next stop.

About

Astronauts on Mars will walk out to study rocks, and must always be able to walk back.

Mission: Vesta is a set of tools to plan those walks with NASA’s own Mars data. We built it for NASA Space Apps 2026, on a 2 × 2 km slice of lower Mount Sharp in Gale crater, where the Curiosity rover drilled.

Why: there is no map of Mars detailed enough for a single walk.

~1.5 m

The smallest rock orbit reliably sees. The ones that trip a walker are smaller.

4–22 min

For a message to reach Earth. Mission control can’t guide each step.

~8 h

Of oxygen in a spacesuit. Every step out has to leave enough to walk back.

So we build that map: start from NASA’s orbit data, scout the route from the air, then walk with the way home always in view.

Method · our process

Look first, then walk.

Orbit sees the whole site, but not the small rocks. A scout helicopter sees them, but only along its path. So each step tells the next one where to look.

  1. Choose stops. Pick at least three science stops on the orbit map, such as Curiosity’s drill sites.
  2. Plan the route. The app finds the path that costs the least energy. Slopes over 20° are walls.
  3. Scout. A simulated helicopter flies the route first and finds rocks orbit can’t see. The route is planned again around them.
  4. Walk. A ring shows how far you can go and still get home with the reserve untouched. The app warns at the point of no return.

Vesta Scout runs these four steps as a guided mission. Vesta Vision is the map underneath: every layer, in 3D, down to eye level.

Vesta Scout

Look before you walk.

Pick science stops from the habitat. A simulated scout helicopter, like NASA’s Ingenuity, flies the route first and finds the rocks orbit can’t see. Then the walk is planned around them.

Planned from orbit: 908 m, 435 kJ. It crosses 2 rock cells the scout later found.

Open Vesta Scout
HiRISE map of lower Mount Sharp: the route planned from orbit, dashed red, runs through two rocks the scout found. The same map: the scouted route, white, bends around those rocks.
Habitat (H) and three stops, lower Mount Sharp, Gale crater. Red dashes: the route from orbit. White: after scouting. Red squares: rocks the scout found. NASA MRO HiRISE.
Read the method
Route
Least-energy path over 8 m cells of the HiRISE terrain. Each step costs energy from the Pandolf walking equation, scaled to Mars gravity (0.68, an estimate). Slopes over 20° are walls.
Scout
The flight is simulated. The rocks it “finds” come from the stored rock layer: a roughness estimate from the terrain, plus rocks the ROSCO detector proposed on a HiRISE image and a person confirmed (23 of 25). Coverage is not complete, so we claim no hit rate.
Walking back
Before the walk, the energy to get home is worked out for every cell. On the walk the app only looks it up, and warns at the point of no return.

Vesta Vision Early version

Read the ground.

The site as a 3D map: NASA’s 1 m terrain with HiRISE colour, and the layers behind every route.

Stand 1.7 m above the ground and walk. A red guide line follows the planned route, and the label names the next stop and how far it is.

Open Vesta Vision
Eye level on lower Mount Sharp: a red guide line runs to stop 1, labelled Next: Altadena drill site, 345 m. 3D view of lower Mount Sharp in HiRISE colour, with the habitat, five stops and a route. The same view coloured by slope: green flats, dark steep ridges, red ticks on high steps. The same view coloured by energy to walk home: dark near the habitat, paler further away.
Lower Mount Sharp, 2 × 2 km. NASA MRO HiRISE colour and stereo terrain (USGS MSL Gale mosaic and DEM).
Read the method
Terrain
USGS’s 1 m Gale terrain model from HiRISE stereo, averaged to 8 m cells for planning. Slope and steps come from it.
Ground at each stop
A ground classifier (MobileNetV3-small, trained on NASA/JPL AI4Mars labels) runs in the browser. 92–94% on held-out Curiosity patches and 73% on fresh Perseverance patches (the first version scored 41% there).
Energy home
Dijkstra’s shortest path, run once from the habitat over the Pandolf step cost. Red cells have no way back.

Docs

How it works, step by step.

No background needed. Each step says what we do, how, and what it gives the astronaut. Every number comes from the code or a file in our repository; estimates are marked.

  1. From NASA data to a map

    We take NASA’s MRO HiRISE images (25 cm per pixel) and the USGS 1 m terrain model made from HiRISE stereo of lower Mount Sharp, Gale crater. We cut a 2 × 2 km slice into 8 m squares, 256 × 256 of them. Each square knows its height, its slope and its ground type.

    For the astronaut: one map of the site where every square has a known slope and cost.

  2. What a suit can cross

    Slopes over 20° (NASA xEVA testing) and steps over 0.3 m (NASA study ICES-2025-294) are walls. Rocks count as walls once someone has seen them.

    For the astronaut: the route never asks them to climb what a suit can’t.

  3. The route

    Each step costs energy from the Pandolf walking equation, scaled to Mars gravity (0.68, an estimate) and harder on sand. Dijkstra’s algorithm finds the cheapest path between every pair of stops, then the app tries every order of the stops (Held-Karp) and keeps the one that fits the energy left after the reserve: 1,266 kJ usable, with 60% of the suit budget kept back.

    For the astronaut: the least tiring route to as many stops as the suit allows.

  4. The scout, and the Canadian Traveller Problem

    Planning a walk when some squares may hide rocks orbit can’t see is a known problem in computer science: the Canadian Traveller Problem (CTP) with remote sensing. The map is known, some links may be blocked, and you can pay to look before you travel. Our way to look is a helicopter. It flies the planned route first (up to 1,000 m a flight, seeing 15 m on each side), every rock it finds becomes a wall, and the route is planned again after each landing. The flights are simulated; the rocks they reveal are the stored rock layer.

    We also built a value-of-information planner that picks where to fly by the walking cost it expects to save over 12 sampled versions of the site (flag voi, off). On Gale it chose no flight and left 6 rock cells on the route, where flying the route left 0, so it is not the default.

    For the astronaut: on Gale the route from orbit crosses 6 rock cells; after scouting it crosses none, for 621.4 → 629.7 kJ. The detour is decided in the habitat, not mid-walk.

  5. The way home from everywhere

    Before the walk, Dijkstra’s algorithm runs once from the habitat and gives every square the least energy needed to walk home: the cost-to-go map. On the walk the app does no route search; it looks the number up. The walk-back ring is every square where cost to go is at most the energy left minus the reserve. Oxygen left is energy left ÷ 20.1 kJ per litre.

    For the astronaut: at every step they see how far they can still go and get home with the reserve untouched, and the app warns at the point of no return.

How the models were trained.

Ground classifier

Data: NASA/JPL AI4Mars, Curiosity and Perseverance navigation-camera images labelled by people. 2,037 training patches of 256 × 256 pixels (1,136 Curiosity, 901 Perseverance) in four classes: soil, bedrock, sand, big rock.

Model: the photo turned grey (Curiosity's camera is greyscale), MobileNetV3-small image features (pretrained, frozen) and a small linear layer we trained: 200 epochs, Adam. It runs in the browser.

Result: 92% and 94% on two held-out Curiosity tests (159 and 302 patches); 73% on 219 Perseverance patches from sols it never saw, where the first version scored 41%.

For the astronaut: names the ground at a stop from a photo. It does not steer the route: it is not yet tested on helicopter or phone photos.

Rocks from orbit images

Data: one HiRISE image of the site.

Model: NASA/JPL’s ROSCO rock detector proposed rocks; a person checked each one.

Result: 23 of 25 proposals confirmed. Coverage is not complete, so we claim no hit rate.

For the astronaut: real rocks in the layer the scout reveals.

Chance of hidden rocks

Data: our stored rock layer (a terrain roughness estimate plus the reviewed rocks), west half for fitting.

Model: logistic regression on two orbit features: slope and 8 m relief.

Result: AUC 0.73 on a development check. Not calibrated: only 3 of 22 reviewed rock cells could be tested.

For the astronaut: an experimental hint of where to look first (flag belief, off).

Sources and links.

NASA data on the map

  • MRO HiRISE: 25 cm colour images (USGS MSL Gale mosaic) and the 1 m terrain model from HiRISE stereo (USGS MSL Gale DEM).
  • NASA/JPL ROSCO: rocks proposed on a HiRISE image, checked by a person.
  • NASA AI4Mars: labelled rover images that train the ground classifier.
  • Curiosity: its drill sites as stops, and ChemCam, DAN, REMS and RAD readings near each one.
  • Viking Orbiter Mars globe (NASA/JPL-Caltech), at the top of this page.

Where the rules come from

Real or simulated

  • Real: terrain, slope and images, the Curiosity stops, the walking-energy formula, all the route and safety maths.
  • Estimates, marked: ground type, the rock layer, the energy budget and the walking-energy factors.
  • Simulated: the helicopter flights and the astronaut (the app in a desktop browser stands in for the suit display; it is built for a landscape view, so a phone may not give the best experience).

Go deeper

NASA Space Apps 2026. Data: NASA MRO HiRISE (NASA/JPL-Caltech/University of Arizona), USGS MSL Gale mosaic and terrain model, NASA/JPL ROSCO and AI4Mars, NASA Curiosity, Viking Orbiter Mars globe (NASA/JPL-Caltech).