Scientific exploration of challenging planetary analog environments with a team of legged robots

In short: the places on the Moon that scientists most want to study are steep, loose and rocky, which is exactly where wheeled rovers can’t safely drive.

Robots are how we explore the solar system, and increasingly how we look for resources there. Space agencies and companies have committed to returning to the Moon, this time aiming for a lasting human presence. NASA’s Artemis program is headed for the lunar south pole, and one of its first missions, the VIPER rover, was planned to visit permanently shadowed regions: cold hollows that have not seen sunlight for millions of years and are thought to be rich in ice and other volatiles.

Many of the most valuable targets sit in hard places: volcanic vents and channels (rilles), caves, unusual “irregular patches” in the dark lava plains, and fresh impact craters. The south-polar terrain around them has steep slopes, fields of boulders thrown out by impacts, and soil that may behave in unexpected ways.

Wheels are the bottleneck

Every rover that has driven on the Moon or Mars, from Lunokhod 1 in 1970 to Perseverance, rolls on wheels. Wheels are well proven on fairly flat ground but reach their limits on steep slopes, loose granular soil and rough, blocky terrain. On Mars, Spirit was lost in unusually loose soil and Opportunity was temporarily stuck in a dune. On the Moon, the Apollo 15 astronauts had to lift their rover out of loose soil by hand, and Lunokhod 2’s wheels sank more than 20 cm near Le Monnier crater. The team behind China’s Yutu-2 rover said craters would be of great scientific interest, but they don’t drive into them because the risk of getting stuck is too high. In the paper’s words, this limitation keeps current missions away from high-priority targets.

There is a second limit: a single robot carries a single set of instruments and explores at a single robot’s pace. And on the Moon, every command and every reply takes seconds to travel.

A real-world test: the Space Resources Challenge

The European Space Agency (ESA) and the European Space Resources Innovation Centre (ESRIC) set up the Space Resources Challenge in 2021 to push lunar prospecting robots forward. Teams had to search an unknown Moon-like site for resource-enriched areas: spots with minerals such as ilmenite, rutile and titanium dioxide that could be used on the Moon itself (in-situ resource utilization). The conditions copied the lunar south pole: unknown terrain, loose soil, a low sun casting long, harsh shadows, a 5.0 s round-trip delay on every message, and moments when the link dropped out completely. The challenge inspired this work and was one of its two big field tests.

What existed before

• Legged robots on Earth already walk robustly over mud, gravel, snow, vegetation and sand, using controllers learned in simulation.
• Legged robots for space, including this lab’s own, had shown dynamic walking on steep planetary soil and in low gravity. But to be useful they must do more than walk: carry instruments, measure, take samples. At the first round of the Space Resources Challenge the authors used a single legged robot with instruments fixed to its body.
• Robot teams. All top teams in the DARPA Subterranean Challenge (2021) used mixed robot teams, including the lab’s team CERBERUS, which this work builds on. Those teams differed in how they moved, not in what they could measure, and had no instruments to place or long delays to cope with.
• Planetary robot teams have been tested on Earth: the German Aerospace Center (DLR) sent a drone and two wheeled rovers onto Mount Etna with a high level of autonomy, but wheels limited where they could go; DFKI paired a wheeled and a legged robot to collect samples; another team ran a remote-controlled sample-return test, but checked skills like sampling separately rather than in a full mission. NASA JPL has sent several Spot robots into Mars-like caves (not yet published in detail). On Mars itself, the Ingenuity helicopter scouts for the Perseverance rover.

Where this paper fits: a team of legged robots that differ in their science skills, tested end to end in realistic missions with a long delay, on terrain where wheels struggle.

In short: on loose, steep soil wheels slip and dig in. The authors checked that a walking robot doesn’t, up to the 25° limit of a tilting Mars-sand test rig.

Why would legs do better? A wheel has to roll through whatever is under it. On loose sand it spins, digs a rut and can sink until the vehicle is stuck.

A walking robot instead places each foot and pushes off, so a slipping foot does not turn into a rut that swallows the vehicle. To test this, the authors took the Scout and its existing learned walking controller (a neural network trained by trial and error in simulation; Miki et al., 2022) to the ExoMars rover testbed at Beyond Gravity: a 6 m × 6 m box of ESA’s ES-4 Mars soil simulant that tilts up to 25° in 0.1° steps.

• Up and down the 25° maximum on ES-4 soil and on limestone “bedrock”, three runs each, with no fall and no stall.
• Top speed on the 25° slope: 0.7 m/s.
• At 10°, it crossed very loose hills with deep foot sinkage and high steps between soil and bedrock.
• In the field missions it climbed granular slopes of up to 20° and a crater rim, and walked 358 m of granular terrain at the Space Resources Challenge. There was no locomotion failure on any robot in any test or mission.

What this means: the legged robot is roughly 30–40× faster than today’s planetary rovers on flat ground and handles steeper loose slopes; 25° was simply the limit of the test rig, not of the robot.

In short: instead of one robot that does everything, three robots split the work the way a field geologist does: look around, look closer, then study a few rocks in detail.

All three robots are ANYmal quadrupeds from ANYbotics: 50 kg, 15 kg payload, about 90 minutes of continuous walking. The Scout is an ANYmal C (2019); the Hybrid and the Scientist are ANYmal D (2021). They share the walking, mapping, localization and navigation software, and differ in what they carry. The split follows how a field geologist works: first get an overview, then look at many candidates from a distance, then study a few in detail.

What this means: the instruments get closer to the rock as you go down the team: images from metres away, Raman spectra from 2 m or 0.7 m, and microscope images in contact.

Two terms recur below. Remote measurement tasks are images from the CTX cameras, aimed at a 3-D point. In-situ measurement tasks are MIRA and MICRO readings, which need a full 6-D instrument pose (position and orientation) on the target.

In short: messages to the robots arrive seconds late, so the system sends small maps and gives goals (“go there”, “photograph that”) instead of steering the robots.

Lunar operations have a round-trip delay of seconds, and the link can drop out. Both missions simulated this with a delay box that held every packet for a 5.0 s round trip. That rules out joystick driving and streaming big maps, so the system is built around two ideas: send little data, and send goals instead of commands.

Two maps: one to steer by, one to keep

• Localization: LiDAR SLAM (simultaneous localization and mapping: building the map and finding yourself in it at the same time; a modified CompSLAM) with inertial data as a prior; when the terrain gives the LiDAR too little structure in some direction, it falls back on the prior in that direction.
• Dense map, kept on the robot: an octree (a 3-D grid that only stores occupied space) with 30 mm voxels (3-D pixels), filtered (points above 2 m clipped, points with fewer than five neighbours within 200 mm removed, points on other robots removed) and coloured from the navigation cameras. It is fetched after the mission and turned into a Poisson mesh (depth 12, up to five minutes on an i9 desktop).
• Light mesh, sent during the mission: the robot downsamples to 150 mm voxels within 9 m, compresses with Draco (3 MB → 250 kB), and sends only when there is something new. Mission control meshes it (Poisson, depth 8, at most 0.25 Hz), drops weakly supported triangles and fuses the patch into the map.
• Elevation map for walking: an 8 m × 8 m, 40 mm grid around each robot, updated on the GPU, feeds the local planner.

Boulders flagged automatically

A Mask R-CNN network (a standard neural network that outlines objects in images), fine-tuned on a few hundred images the team collected at the first SRC field trial in dark, lunar-like conditions, runs at mission control on navigation and zoom-camera images. It outlines each boulder with a box, mask and confidence, which saved the scientists a lot of looking.

Goals, not joysticks

Operators send high-level goals: a 2-D navigation goal, a 3-D point to image, or a 6-D instrument pose. A behavior tree (a flowchart of steps, with fallbacks for when a step fails) on each robot runs the task to the end, so each task costs one operator interaction. The trees share subtrees across robots (navigation, inspection, measurement); the Scientist’s is the largest, with 76 leaf nodes. A sampling-based local planner works on the elevation map, modelling the robot as foot reachability volumes plus a torso collision volume rather than labelling terrain as good or bad.

Networking: Rajant mesh radios at 5.8 GHz, one ROS master per robot and station, and Nimbro Network between them over UDP, because TCP handshakes stall with a 5 s delay. Data products were kept small so that a single lost packet doesn’t ruin a whole image or map. Incoming traffic at mission control stayed below 3.5 MB/s (Fig. S8).

In short: a heavy arm on its back pushes a walking robot around; training the walking controller with the arm included made it walk more naturally and use 15 % less power.

All three robots walk with a reinforcement-learning policy (a neural-network controller trained by trial and error in simulation) from the perceptive-locomotion pipeline of Miki et al. (2022). It is trained in two stages in simulation: a teacher that sees privileged ground truth, then a student that copies the teacher from noisy, realistic sensing. The Scout uses that controller unchanged.

The Hybrid and the Scientist are different. They carry heavy payloads, and the Scientist’s arm moves while it walks and pushes the base around. The old controller did not know about any of this and compensated by walking with a low body, which needs large knee torques and wastes energy. The new policy changes the training in three ways:

1. Put the arm in the simulation and in the observation. The arm is simulated with its real mass and moves to random targets under its own PD controller; the policy (which only commands the legs) observes all 18 joints (12 leg + 6 arm), with position and velocity histories. This replaces an earlier approach that fed in a predicted wrench from the arm’s MPC.
2. Randomize payload mass and tell the teacher. The extra mass is drawn from U(−2, 7) kg and given to the teacher as privileged information.
3. Penalize torque more in the reward, to favour efficient gaits.

What this means: the gain is largest when standing and walking are mixed, as in a real measurement mission; even while walking, the new policy uses noticeably less power (Mann–Whitney U test, P < 0.001).

Note: the supplementary text lists the walking-segment means as “808 W and 749 W using our controller and the baseline, respectively”, but Fig. S9 and the surrounding text show the new controller is the lower one. The table follows the figure.

In short: a scientist clicks on a rock once; the robot walks up, aims a mineral-reading laser, and presses a microscope gently against the rock on its own.

Once a target is chosen, the Scientist does the close-up work by itself:

1. Pick. Near the target, the operator clicks the spot in the Realsense grey-scale infrared stream (it works in bad light). A 6-DoF interactive marker appears and can be adjusted. The operator chooses MIRA, MICRO or both.
2. Approach. The robot walks up and parks its base 1 m from the target.
3. MIRA. The arm points the Raman laser at the target. A control loop (after Abi-Farraj et al.) drives the alignment error below 0.01 m at the ideal stand-off of 0.7 m. MIRA auto-focuses, measures, and matches the spectrum against a library of 355 reference spectra of lunar-relevant minerals (oxides such as ilmenite and rutile, olivine, pyroxenes, feldspars, regolith simulant).
4. MICRO, stage 1. The arm moves MICRO to 100 mm in front of the target pose.
5. MICRO, stage 2. A PID loop moves it along its axis using the time-of-flight sensor until the reading is 5 mm, which gently squeezes the foam ring onto the rock. This works even if the clicked pose was not exactly on the surface. MICRO then focuses and records images under white, red, green, blue, UV (395 nm) and near-infrared (940 nm) light.

If anything goes wrong, the arm returns to its default pose and the robot tells the operator. The arm itself is controlled by model-predictive control (MPC: it re-plans the arm’s motion a short time ahead, many times a second; OCS2 library) that treats the robot as a floating-base manipulator; while walking it holds the arm in a low, compact pose, and during measurements it tracks the tool pose. Arm and legs are controlled separately: MPC for the arm, the RL policy for the legs.

In short: the robots’ instruments overlap on purpose, so a teammate can take over when a robot or instrument breaks. At the quarry, that is what saved the mission.

Each robot has a main role, but the payloads overlap on purpose, and any mission-control station can command any robot. So when a robot or an instrument fails, the operators reassign its tasks. Table S1 lists what is still possible for every combination of offline robots:

What this means: losing any single robot still leaves four of the five measurement types (◦ = limited quality: the Hybrid’s base-mounted MIRA can’t be pointed as precisely as the arm’s). A single robot carrying everything would lose all five.

This was tested for real at the quarry. The Scout’s CTX-FW and the Scientist’s MIRA malfunctioned. The team sent the Hybrid in early to take over the Scout’s camera work and the Scientist’s Raman measurements, and brought in the Scientist once enough targets were queued for its microscope. Every mission goal was met except the filter-wheel images: six of seven.

In short: two full practice missions under Moon-like conditions: most of an unknown site mapped, targets found and measured within an hour or two, and no robot ever fell.

Both end-to-end missions had five people in the control room: at each of two stations, one robot operator and one planetary scientist, plus a supervisor. One station ran the Scout and the Hybrid, the other the Scientist. Only one station commanded a robot at a time, so commands never conflicted.

Space Resources Challenge, Luxembourg (Sept 2022)

The ESA/ESRIC challenge asked teams to prospect an unknown 1800 m² lunar analog: coarse basalt with fine dust, a powerful lamp at a low angle in one corner (south-pole lighting), a 5 s round trip, one scheduled and one random loss of signal, and five hours for the mission plus the report. The Scout went first. After the first loss of signal the team had enough targets to send the Scientist to the candidate resource areas; the Hybrid joined at 76 min for boulders. For 20 minutes all three robots worked at once.

What this means: because the Scout had already mapped the area, the Scientist spent its time measuring instead of exploring; that division of labour is the main argument for a specialized team.

Quarry, Neuheim, Switzerland

An active gravel quarry with poorly sorted sediment, metre-sized boulders and a headwall up to about 20°. The test ran at night, lit by a single 180 W LED at about 87° to mimic south-pole shadows, again with a 5 s delay. The team scattered patches of basalt, ilmenite, rutile and titanium dioxide as mock resource areas. Goals: map at least 1000 m², find at least five targets, and measure at least five targets with each instrument.

What this means: with two instruments broken, redundancy turned what would have been a failed single-robot mission into a nearly complete one.

When the link drops

During a loss of signal each robot finishes its current task, then runs a fixed routine. The Scout and Hybrid take a panorama, photograph their own footprints (soil mechanics), photograph each other and the Scientist (hardware check), and the Scout photographs the colour card for later colour correction. Everything is sent once the link is back.

In short: it worked in practice missions on Earth, but people still make every decision about what to study next, and the hardware is not ready for space yet.

• Instruments. The payload mix was set by budget and time. Other SRC teams found resource areas better with X-ray fluorescence (XRF); adding XRF or LIBS, or returning samples to a lander, would improve the science.
• The 5 s delay hurts. TCP is unusable, UDP loses packets, and every operator interaction costs mission time. Better protocols and compression would help.
• Humans are the bottleneck. The robots run single tasks alone, but people still choose and assign every task. Automatic target selection and task allocation would need machines to interpret images and spectra, which is still open. More robot-to-robot communication (shared maps, task hand-offs without mission control) is also missing.
• More autonomy for Mars. Longer delays need multi-goal planning, automatic target prioritization and skill-aware task allocation, which would also let the team grow without more operators.
• More variety. The robots differ in science skills, not in mobility. Drones for fast mapping and wheeled robots for easy, long-range targets could complement the legged ones.
• Space hardware. Power and thermal design for legged robots look feasible, but there are no space-grade processors fast enough for this mapping, navigation and locomotion stack, and the LiDARs are not space-qualified (solid-state LiDARs may change that).