Marrow
Uncrewed ground vehicle · Logistics · Casualty evacuation
Fourteen hundred pounds through terrain a ground robot normally refuses — and a casualty deck that stays level to two degrees while it does it.
Design · no prototype
Mission
The squad's terrain is the design case. The trail is not.
Carry 635 kg — 1,400 lb — of a dismounted squad's load through the ground the squad actually walks in, and bring casualties back out of it on a deck that stays level.
The load problem is old and it is not solved. A rifleman carries 45–60 kg. A tracked mule fixes that on a trail and gives up at the first 600 mm step, the first scree slope, the first blowdown — and then four soldiers spend twenty minutes recovering it. MARROW exists to take the squad's water, ammunition, batteries, mortar rounds and sensors into terrain that stops a wheeled or tracked vehicle, and to bring a wounded soldier back without occupying four riflemen on a litter carry.
It is a wheeled-leg hybrid: four independently articulated three-degree-of- freedom legs, each terminating in a driven wheel. It rolls when it can and walks when it cannot. The whole product is that one architectural choice and the two things it buys — terrain access, and an actively levelled deck.
Key figures
Locomotion
Borrowed, not re-specified
MARROW does not have its own locomotion program. It has the Fantom Robotics one, retargeted.
The inheritance
The gait controller, contact-state estimator, whole-body QP solver, terrain-cost network and sim-to-real training pipeline are the Fantom Robotics locomotion stack, retargeted from the bipedal humanoid program to a four-limb wheeled-leg morphology.
That program is specified elsewhere and this page does not re-spec it. Two things about the retarget are worth stating here.
One: it is real work, not a recompile. A biped controls a two-contact system with a narrow support polygon and a high centre of mass. MARROW is a four-contact system with a wide polygon and a low centre of mass carrying a 635 kg live payload whose own centre of mass moves. The whole-body QP is the same solver. The contact schedule, the inertial model and the terrain cost are not.
Two: what comes free is the thing that matters most. Because the four legs are individually height- and force-controlled, the cargo deck is actively levelled to ±2° in pitch and roll independent of the ground. A tracked vehicle cannot do this at any price, and it is the difference between a casualty being carried and a casualty being rolled down a hillside on a stretcher.
Every fielded MARROW streams trajectories back into REVENANT FORGE and re-trains the fleet against the same simulator as the humanoid program. A UGV program that writes its own controller from scratch is a five-year program. This is a retarget.
| Road speed | Tracked mule 8 km/h · legged robot 5 km/h · MARROW 32 km/h |
|---|---|
| Cost of transport, improved surface | Tracked low · legged 4–5 × higher · MARROW low (it rolls) |
| Vertical step | Tracked ~250 mm · legged 700 mm+ · MARROW 610 mm |
| Deck levelling | Tracked none · legged yes · MARROW ±2°, active, under load |
| Actuator count | Tracked 2 drive motors · MARROW 12 leg actuators + 4 hub drives |
| Deep saturated clay | Tracked wins. Legged loses badly. MARROW loses. See the risks. |
| The price of the hybrid | 78 kg of hip and knee actuator a tracked vehicle does not carry, plus twelve high-torque actuators as wear items instead of two motors. What it buys is terrain access and a level deck. That is a defensible trade for a squad-support vehicle and a bad one for a supply-route hauler — and we are not building a supply-route hauler. |
Specification
Parameters
Design targets, not measured results, unless the line says otherwise.
| Mass, empty | 780 kg |
|---|---|
| Payload | 635 kg (1,400 lb) |
| Gross vehicle weight | 1,415 kg |
| Dimensions, stowed | 2.28 × 1.12 × 1.05 m — narrower than an ATV, fits a trail |
| Transport | Two internal in a CH-47F; sling two; one in a HEMTT bed or a Bandvagn |
| Ground clearance | 180–720 mm, actively variable by leg extension |
| Deck height | 380 mm squatted for casualty loading, 900 mm standing |
| Deck levelling | ±2° pitch and roll, active, independent of ground |
| Vertical step | 610 mm at walking gait |
| Gap crossing | 900 mm |
| Side slope / gradient | 35° / 60 % (31°) |
| Fording | 800 mm |
| Environmental | −40 °C to +55 °C, IP67, MIL-STD-810H |
| Unit cost target | $410,000 at 400/yr from FORGE-1 |
| Improved surface, rolling | 32 km/h |
|---|---|
| Off-road, rolling | 9 km/h |
| Walking gait, broken ground | 3.4 km/h |
| Stair / rubble crawl | 1.1 km/h |
| Powertrain | Series hybrid — 22 kW JP-8 genset + 14 kWh LFP. The genset is a range extender and a charger, not the drive. |
| Fuel | 68 L |
| Range, improved surface | 520 km hybrid |
| Range, cross-country at a squad's pace | 190 km hybrid, assuming ~15 % of the distance in walking gait |
| Silent move, battery only | 4.5 h / ~40 km at 9 km/h. This is the operationally decisive number, not the 520 km. |
| Silent watch, parked and sensing | 58 h at 240 W |
| Acoustic, battery, walking | 42 dBA at 10 m — no exhaust plume, no infrared signature above ambient |
| Acoustic, genset running | 61 dBA at 10 m. The genset is the signature. |
| Sensors | 32-beam lidar, 4 × global-shutter stereo pairs (360°), 2 × LWIR, tactical-grade IMU |
| Optional mast | VIGIL-M sensor head on a 4.6 m mast — MARROW becomes a mobile surveillance tower |
| Navigation drift, GNSS-denied | < 0.4 % of distance travelled. A 5 km patrol closes with under 20 m of error, then re-registers on the terrain model. |
| Degraded mode | A failed leg actuator locks that leg extended; the vehicle drives home on three at 4 km/h with 40 % payload |
| Litters | Two NATO litters, or one litter plus 320 kg of Class V |
|---|---|
| Loading | Deck lowers to 380 mm — a two-person litter carry, not a lift |
| Levelling | ±2° pitch and roll for the entire movement. This is the feature. |
| Medical provisions | 500 W outlet, O₂ bottle mounts, vacuum-splint interface, litter-strap hardpoints |
| Rewarming | Deck holds 34 °C from the genset coolant loop — which matters more in a cold-weather casualty than most equipment lists admit |
| What it is not | A litter carrier, not an ambulance. No medic aboard, no monitoring, no intervention. It moves a stabilised casualty from where he was hit to where the medic is. It does not replace the medic and we will not let it be sold as though it does. |
| Follow | Holds 3–40 m standoff on a UWB/optical tag in the squad leader's kit. Follows the leader's track, not his straight-line position. |
|---|---|
| Breadcrumb | Retraces its own logged path to the LZ or patrol base with no operator and no network |
| Waypoint | Autonomous route over the resident terrain model and route graph |
| Teleop | Direct control from a VEIL HUD or a handheld, 9 ms cue latency, to 1.2 km on MESH |
| Localisation | Visual-inertial odometry on four stereo pairs fused with 32-beam lidar and leg/wheel odometry. The contact-state estimator knows when a wheel is slipping, which is a better odometry source than wheel encoders alone. |
| Record | Every autonomous manoeuvre LEDGER-signed. A family will eventually ask what the machine did with a wounded soldier aboard. There will be an answer. |
Follow-the-squad
The behaviour that matters is that it refuses
MARROW does not follow the leader's footsteps. It plans its own traversable path inside a corridor around the leader's track, because a 1.12 m wide, 1,415 kg machine cannot go where a rifleman goes.
It will decline a gap, route around a blowdown, and take a longer line up a slope. If the corridor contains no traversable solution it stops, says so on the mesh and in the squad leader's VEIL, and holds until it is re-tasked or the corridor widens.
The most common field failure of a follow-me UGV is that it follows. It goes into the creek the squad hopped across, or onto the ledge the squad scrambled up, and then four soldiers spend twenty minutes recovering a machine that was supposed to save them work. A ground robot that refuses and explains itself is more useful than one that tries.
Concept of operations
Six phases
-
Phase 01 — Load
Four minutes at the patrol base
Deck squats to 380 mm. The squad loads 635 kg. Load cells report the distribution and the leg controller re-derives its inertial model before it takes a step.
-
Phase 02 — Move out, silent
Battery only, 3–40 m behind the leader
Rolling where it can, walking where it must, refusing where it should. No operator, no tether, no line of sight requirement.
-
Phase 03 — Sensor haul
The squad's organic tower
At the objective MARROW parks, raises the optional 4.6 m VIGIL-M mast and becomes the squad's surveillance node and mesh relay — 58 hours of silent watch on the battery.
-
Phase 04 — Casualty
Four riflemen keep their rifles
Deck clears, squats, two litters load, deck levels. The squad moves at its own pace instead of at a litter-carry pace, and nobody's hands are full.
-
Phase 05 — Return
Breadcrumb, unescorted
MARROW retraces its own logged path to the LZ with no escort and no network, carrying the casualty. The squad does not have to break to escort it.
-
Phase 06 — Account
Every route decision, signed
The whole movement — every autonomous route choice, every refusal, every manoeuvre with a casualty aboard — is in LEDGER. That record is the reason this feature is fieldable at all.
Engineering rationale
Why it is built this way
Wheels on legs, not tracks
Because the squad's terrain is the design case and the trail is not. The tradeoff: 78 kg of actuator, twelve wear items instead of two, and a 37 % unit-cost premium over a tracked mule. If the customer's terrain is a trail, they should buy the tracked mule and we will say so.
The deck levels because the legs already do
This is the one place where a locomotion stack built for a walking humanoid pays a dividend no tracked vehicle can match. The tradeoff: the levelling authority is bounded by leg travel, so on ground steeper than 35° it saturates and the deck starts to follow the terrain.
Series hybrid, battery-primary
The genset is a range extender and a charger. Silent move and silent watch are the operationally decisive modes. The tradeoff: a series hybrid is heavier and less efficient at sustained road speed than a direct mechanical drive.
It refuses, and it says why
Autonomy that never declines a task is autonomy that has to be recovered. The tradeoff: a machine that stops and asks is a machine that occasionally stops when a human would have pushed through — and the first time it does that in front of a customer, it looks timid rather than correct.
Fantom Robotics, not a bespoke controller
The locomotion stack is trained in REVENANT FORGE against the same simulator as the humanoid program, and every fielded vehicle feeds it. The tradeoff: MARROW's schedule is coupled to another program's, which is a real dependency and not a free lunch.
Signed CASEVAC autonomy
A machine carrying a wounded soldier, unescorted, making its own route decisions, will be scrutinised the first time anything goes wrong. The LEDGER chain is what makes it defensible. The tradeoff: it is the highest-scrutiny feature in the land portfolio and it will be the last one released.
Program candor
Open engineering risks
- Actuator life is the schedule risk, not the software. Twelve high-torque actuators carrying 1,415 kg through impact loading. Design life is 4,000 h to first overhaul and the output stage is the wear item. A tracked mule has two motors and no analogous consumable. This is where the program will slip and where the sustainment cost lives.
- Deep mud. Below roughly 200 kPa of soil bearing capacity — saturated clay, silt, a rice paddy, a churned tank trail in spring — tracks win and legs lose. MARROW's answer is a wide-tyre mode and a gait that reduces peak ground pressure to 68 kPa. That helps. It does not win. We do not beat a tracked vehicle in soft ground and we are not going to claim we do.
- Cost. $410,000 against roughly $300,000 for a tracked mule of the same payload class. The justification is terrain access and the level deck. It is a real premium and it needs a real reason on the customer's ground.
- The genset is the signature. Infrared and acoustic. Battery-only is 4.5 h moving. A patrol longer than that either accepts the signature or plans a charging halt.
- A leg failure is a mission abort, not a breakdown. The degraded mode works — three legs, 4 km/h, 40 % payload — but 4 km/h with a casualty aboard is a long afternoon.
- CASEVAC autonomy is the political risk. The LEDGER chain is the mitigation. It is not a guarantee, and this feature will be released last and deliberately.
Related
What it carries and what tasks it
Contact
Send us your terrain, not your requirement
Give us the actual ground — the step heights, the slopes, the soil, the patrol lengths and the resupply interval. We will tell you where the wheeled-leg trade pays for itself and where a tracked vehicle is the correct answer.
Every figure on this page is a design target derived from first-principles analysis, not a measured result. No MARROW has been built. Where a number here is later measured and comes out worse, we will publish the measured number. No export-controlled information is published on this site.