Large-Scale Autonomous Gas Monitoring for Volcanic Environments: A Legged Robot on Mount Etna

In short: changes in volcanic gas can warn of an eruption, but the best place to measure it is exactly where it is most dangerous for people.

About 29 million people live within 10 km of an active volcano, and between 2010 and 2022 eruptions killed more than 1,000 people. Scientists watch volcanoes with networks of sensors that track three things: the ground swelling, earthquakes, and the gas coming out. These signals often change before an eruption, on timescales from years down to minutes.

Gas can be measured from far away or on the spot (“in situ”); the two complement each other. This paper is about on-the-spot measurement, which tells scientists in real time how much of each gas is present and what that says about the magma below.

On-the-spot measurement means bringing an instrument to fumaroles, small openings and cracks that leak gas and steam, often clustered near active vents. Those places are dangerous: the ground is unstable, the air holds toxic sulfur dioxide (SO₂) and hydrogen sulfide (H₂S), and sudden steam-driven or explosive eruptions can happen.

Why a robot that moves

The authors list four advantages of a mobile sensor over fixed ones: it can adapt (fixed sensors get destroyed, buried by ash or left behind when the gas finds a new outlet), it can be sent quickly when something new happens, it collects context (camera images, heat, 3-D terrain) together with the gas, and it can cover a large area with one expensive, high-quality instrument.

In short: drones and wheeled rovers each get part of the way; a legged robot that can walk right up to a vent on its own was the missing piece.

Volcanic gas mixes into the surrounding air quickly, so an accurate reading needs the instrument close to the source. The animation below shows the idea: move the sensor a couple of metres away and most of the signal is gone.

Drones are good for the tall plume rising from a crater, because they can fly into it while keeping a safe distance. But much of a volcano’s gas escapes through fumaroles spread across its flanks, close to the ground, where complex terrain and unsteady wind make it hard for a drone to get close.

Ground robots built for volcanoes exist, but they face loose tephra (volcanic gravel and ash), steep slopes, lava ridges and boulder fields. Wheeled and wheeled-legged rovers adapt poorly to uneven ground, so they must move slowly to protect their instruments and avoid rolling over. Most are driven by remote control, and radio links become patchy in rough terrain or during eruptions. To the best of the authors’ knowledge, none had completed a successful on-the-spot gas measurement on an active volcano.

Legged robots have already worked outdoors, from watching wildlife and sampling soil (remote-controlled) to planetary-analog exploration and forest surveys (partly autonomous). But in those systems an operator usually lays out a dense list of waypoints. The step here is to go further with little operator input, and to carry a real analytical instrument.

In short: instead of cheap single-gas sensors, the robot carries a compact mass spectrometer that can tell many gases apart, from dense plumes down to traces.

The robot is ANYmal, a 50 kg four-legged robot from ANYbotics with three motors per leg. It can carry 12 kg, and its 907 Wh battery lasts about an hour at full payload. For seeing, it has six depth cameras, a spinning LiDAR, two wide-angle colour cameras, a 198° high-dynamic-range camera and a GNSS receiver used without correction signals.

Why a mass spectrometer

The usual field sensors each target a few gases (infrared sensors for CO₂, electrochemical cells for SO₂ or H₂S, and so on). They work, but have a limited range, can react to the wrong gas, and degrade in acidic volcanic air. A mass spectrometer instead sorts molecules by their mass (strictly, mass divided by charge), so one instrument reads hydrogen (2 amu), helium (4), CO₂ (44), SO₂ (64) and many more. Its range spans six orders of magnitude, so it does not saturate inside a dense plume and still sees sub-ppm traces downwind.

The instrument is an INFICON Transpector MPH. Air enters through a 1 m stainless-steel capillary that limits the flow. A small diaphragm pump brings the pressure down to about 10⁻³ bar, then a tiny turbomolecular pump down to about 10⁻⁷ bar. There an electron beam turns molecules into ions, and a quadrupole (four rods with oscillating voltages) lets only one mass reach the detector at a time.

What this means: a lab-class instrument fits within the robot’s 12 kg payload and runs off its battery, with a reaction time of a few seconds, fast enough to notice a plume while walking past it.

A dedicated ROS package talks to the instrument, so every gas sample is time-stamped with the robot’s position and shown live to the operator as coloured dots along the path.

In short: the software is a chain from a few clicks on a map down to motor commands, with navigation parts that decide where to go and perception parts that tell it where it is and what the ground is like.

The paper’s architecture has the same two colours as our legend: blue navigation modules (global and local planning, walking) and green perception modules (state estimation, elevation mapping).

Mission planning and the global route

The planning screen shows a map in latitude and longitude. High-resolution satellite images are not always available, so the team used a topographic OpenStreetMap layer (trails and dense contour lines that reveal craters), or drone photos when needed. The operator places targets; a global planner then runs A*, a standard shortest-path search, over a trail network extracted offline from OpenStreetMap. Targets off the network are joined to the nearest trail point by a straight line.

The cost between trail points is simply straight-line distance. Slope and surface would matter, but public elevation models are too coarse (10–30 m cells) to judge them in advance.

The rest of the chain, state estimation, the elevation map, the lookahead waypoint, the local planner and the walking policy, has its own sections below.

In short: to follow a route drawn on a world map, the robot combines satellite positions, laser-scan mapping and leg odometry, because none of them is good enough alone on a volcano.

The estimator builds on Holistic Fusion, an optimisation-based framework that finds the most likely robot trajectory given all sensor data (a “factor graph”). It combines:

• GNSS without real-time corrections: tied to the globe, but noisy (about ±2–3 m on Etna).
• LiDAR SLAM (the degeneracy-aware X-ICP): smooth and consistent, but it slowly drifts over long distances when the ground offers little shape to match, and when slipping feet shake the robot and blur the scans.
• The robot’s own leg-and-IMU odometry: stable over short times.

The authors made two changes for long field runs. First, SLAM poses enter the optimisation as absolute poses rather than step-by-step increments. Second, the optimiser also estimates the transform between the drifting SLAM map and the world frame, T_W,M. That lets it keep a smooth, globally consistent path through GNSS dropouts, quickly re-anchor when satellites return, and model how uncertainty grows with drift.

Finally the route, given in latitude, longitude and altitude, is converted into a local East-North-Up frame anchored at the robot’s start (on the WGS-84 ellipsoid, with the EGM2008 geoid model to get heights above sea level), so the local planner can use it.

In Mission 3 a narrow valley made GNSS worse (about ±6 m) but the SLAM frame kept the pose stable. On the featureless crater rim of Mission 1, SLAM drift and GNSS noise together were misread as a small heading change, rotating the projected path. The scene’s traces are illustrative.

In short: a live 3-D height map scores the ground ahead, a moving waypoint pulls the robot along the route, and a learned walking policy that knows about the heavy payload does the stepping.

Elevation map and traversability

Elevation Mapping CuPy fuses the six depth cameras and the LiDAR on the GPU into a robot-centred map of 12 m × 12 m at 6 cm resolution. A lightweight neural network, trained on a small set of hand-labelled terrain samples, scores each cell’s traversability from 0 to 1. The scores are then smoothed (inpainting) and dilated with a 0.24 m kernel for extra safety margin near rough ground.

The lookahead waypoint

The local planner can only see what is around the robot, so a path manager hands it a lookahead waypoint on the global path, updated at 0.5 Hz. Too far ahead and the robot cuts corners into dead ends; too close and it chases every wiggle and offset in the map path. So the reach depends on how sharply the path bends:

The waypoint itself is an exponentially weighted average of points sampled along that stretch of path, so nearby geometry counts most.

The local planner

The Field Local Planner blends two “force fields”. Cells with traversability below 0.2 count as obstacles, and a signed distance field built from them pushes the robot away. A geodesic distance field, the shortest walkable distance to the waypoint from every cell, pulls it toward the goal.

The walking policy

The legs are driven by a reinforcement-learning policy trained entirely in simulation with a teacher–student scheme (following the lab’s earlier perceptive-locomotion work). The teacher sees hidden facts such as ground friction, extra payload mass and pushes; the student learns to copy it from the robot’s own senses and the elevation map. A standard policy tolerates moderate payloads but walks less efficiently with a heavy, off-centre load, so the team trained an extended policy with the payload’s mass added to the body and its centre of mass randomised.

The policy handles up to 1.5 m/s and 25° slopes; for safety it was capped at 0.8 m/s on Etna.

In short: across a crater rim, a crater descent and a sandy volcanic desert, the robot ran on its own more than 90 % of the time and found five of eight test gas sources.

The team first looked for places with real degassing that they could reach safely. The main summit craters needed a 2.5 km hike with 400 m of climb from the nearest parking, and the Barbagallo craters had a narrow path crowded with tourists. Active degassing was found at the Laghetto crater, but its path was too narrow, steep and sandy for autonomous tests (that site became the remote-controlled run in the next section). The autonomous missions therefore ran at safer sites, with helium bottles as gas sources: helium is non-toxic, non-flammable, inert and almost absent from normal air, so any spike is unambiguous.

• Mission 1, crater rim: a long loop around the Silvestri crater, with few landmarks for localisation.
• Mission 2, crater descent: down into the same crater, over steep slopes and a rocky floor, and back.
• Mission 3, volcanic desert: near the Laghetto crater, with fine sand, rocks, a narrow valley and weak GNSS.

What this means: over roughly ten-minute, few-hundred-metre missions the operator only had to step in for well under a minute in total, and one mission needed no help at all.

Two measures describe autonomy. The autonomy rate is the share of mission time without human input. The robot attention demand (RAD) compares how long each intervention lasts with how long the robot runs between interventions; lower is better, and all three missions stayed below 0.1.

What happened in each mission

Mission 1. Three of five helium sources showed up as clear peaks at the right places; wind along the rim carried the other two plumes away from the robot. Each peak rises steeply and fades slowly, typical for mass spectrometers because some gas lingers in the vacuum chamber.

The four interventions: once the local planner overestimated the gap between two obstacles, twice traversability was misjudged on fine ash, and once the robot turned toward unsafe ground because its heading in the world frame was slightly off. All happened where the projected global path sat several metres off the real trail, for two reasons: the OpenStreetMap trail itself was shifted, and on the weakly structured rim SLAM drifted while GNSS was uncertain by ±2–3 m, which the fusion misread as a small rotation. The waypoints then became unreachable, and the local planner steered into dead ends or poor ground.

Mission 2. Fully autonomous: down a moderately steep, partly loose slope to a boulder field and back up. The elevation map missed some small rocks, so the robot sometimes stepped on stones that shifted, and the walking policy recovered each time. Here the gas lab was switched on only at the target, as it would be for a known degassing area, and the single source gave a clear peak.

Mission 3. Hilly sand with sharp boulders and a narrow valley; the route had to go around a long hill the local planner could not handle. Three interventions: once the policy did not lift a foot high enough past a small hole next to a rock, and twice the robot slipped uphill in fine sand without enough forward speed, so the operator steered by hand. In the valley GNSS uncertainty grew to about ±6 m, but fusion with SLAM held the pose. The valley source was missed on the first pass, probably because wind pushed the plume away from the inlet, and found on the way back.

In short: at real volcanic vents, driven by remote control, the robot’s instrument clearly picked up sulfur dioxide that a handheld sensor carried along the same path did not.

Inside the Laghetto crater the fumaroles sat on a side slope, reached by a narrow trail running along the slope. The trail was compacted, but everything around it was loose, deep sand; even small sideways steps made the feet sink, so autonomous operation was not possible and the robot was driven remotely. Moderate wind kept moving and diluting the plume.

The instrument watched water vapour (18 amu), hydrogen sulfide (34), carbon dioxide (44) and sulfur dioxide (64), with nitrogen (28) as a background reference. The robot walked along the slope, stopping where steam was visible.

• A distinct SO₂ peak appeared when the robot stood right next to a fumarole.
• A smaller rise in SO₂ and CO₂ came as it turned around to head back.
• Water vapour showed no clear increase, probably because the air was already humid.

For comparison, the team carried a handheld miniGAS instrument (INFICON Hapsite Scout, with electrochemical CO₂ and SO₂ sensors) along the same path and held it into visible plumes. Its CO₂ readings marked the same fumarole positions as the robot’s map. Its SO₂ sensor did not register a clear signal, while the robot’s mass spectrometer detected a pronounced SO₂ peak from about 0.5 m away. The handheld’s stronger CO₂ response is explained by being held directly at the source.

What this means: the two instruments agree on where the vents are, and the robot’s instrument sees SO₂ the handheld sensor missed, supporting its use for on-the-spot volcanic gas analysis. The authors describe this as the first successful in-situ volcanic gas measurement by a ground robot on an active volcano, a proof of concept.

In short: the walking was solid; wind, maps that are metres off, dust and very soft sand are what still needs work.

Gas sensing

• Plumes are fleeting, especially in wind. Even a fast, sensitive instrument sometimes missed them, so future systems should predict plumes, for example by spotting gas visually or measuring wind.
• The inlet sits about 0.75 m above the ground, which may be too high for weak, ground-level gas. A deployable inlet that reaches down to the source would help.

Autonomy and mobility

• Planning: map data were coarse and sometimes shifted from the real trail, and large obstacles (like the hill in Mission 3) were invisible in satellite imagery. The authors propose feeding local observations back into the global map, possibly with the robot actively scanning to compare route options.
• Perception and localisation: dust and fine sand stirred up by walking degraded depth cameras and LiDAR; flat, self-similar slopes left scan matching under-constrained; GNSS was poor near crater rims and steep walls. Looks were misleading too: identical-looking sand behaved differently underfoot. Traversability that adapts from what the feet feel is a suggested fix.
• Locomotion: over five days on varied volcanic ground, no falls. But reaching real fumaroles autonomously will need reliable walking on very fine sand on steep slopes, which may call for new foot designs and adaptive contact strategies.

This walkthrough follows the arXiv version of the article (arXiv:2601.07362v2).