DLR Institute of System Dynamics and Control Terramechanics Robotics Locomotion Lab

DLR [5]. CC BY-NC-ND 3.0.

TROLL is a single wheel terramechanics laboratory built around a force-controlled industrial robot rather than a rail carriage. The arm holds the wheel against the soil at a commanded vertical load, imposes a commanded slip, side-slip and tilt, and, through an automated tool changer, swaps the wheel drive unit for a soil preparation tool between runs [1][2]. The consequence is throughput: 115 tests in three days by one person, against several weeks for the same matrix on a conventional manual bench.

Most single wheel rigs elsewhere impose the vertical load with a counterweighted slider on a rail rather than a robot arm. Tohoku University’s testbeds recover rut geometry from LiDAR scans of the surface after each pass [7] and compare wheel performance across simulants of different cohesion [8], and Carnegie Mellon’s soil-imaging rig held its test wheel on a similarly weighted vertical axis to isolate the shear mechanism grousers add [9]. TROLL replaces that slider with a commanded force set point on every axis, trading a fixed load path for one the operator can change between runs without touching the rig, which is the design choice that lets one bench sweep load, slip, side-slip and tilt in a single unattended matrix.

It is at Oberpfaffenhofen-Wessling in Bavaria, and belongs to the Institute of System Dynamics and Control, which is a different DLR institute from the Institute of Space Systems in Bremen that runs the landing rig, and a different one again from the neighboring Institute of Robotics and Mechatronics on the same campus.

Laboratory data

Parameter	Value
Operator	DLR Institute of System Dynamics and Control [3]
Location	Oberpfaffenhofen-Wessling, Germany [5]
Commissioned	First facility paper 2018; “recently developed” in 2019 [1][2]
Type	Robot-based single wheel soil bin, force controlled
Floor area	Not published. Design envelope 6.5 x 6.5 x 3.5 m
Capabilities	Single wheel bench, soil preparation tool, soil bins
Simulant or terrain	DLR-RMCS13 chalk silt; DLR-RMCS14 olivine sand [2][3]
Instrumentation	ATI Omega160 force-torque sensor above the tool changer [1]
Ground truth	Penetrometer at 10 points along the lane, to 0.2 m [2]
Fidelity limits	1 g; usable lane about 3 m of 4 m [1][2]
Access	Institute-run; no published external user program [2]
Cited by	rosalind-franklin

Capabilities

Single wheel bench

Parameter	Value
Working volume	Lane 4 m long, about 3 m usable, wheels to 200 mm wide [1]
Test article limits	Wheel load to 1 kN; wheel radius 80 to 250 mm [1][5]
Vacuum	Not applicable. Ambient pressure, enclosed hall
Temperature	Monitored, not set: 24.5 to 26.5 C [2]
Illumination	Not applicable. No solar simulator
Simulant or terrain	DLR-RMCS13 and DLR-RMCS14, tonnage not published
Slope	Side-slip 0 to 15 degrees instead [1]
Gravity offload	None. Vertical load commanded, 50 N and up [1][2]
Instrumentation	ATI Omega160 force-torque sensor; two Keyence LJ-7300

The machine is a KUKA KR210 R3100 Ultra inside a dust protective jacket, standing between a large fixed soil container behind it and a smaller interchangeable container in front, with an automated tool changer at the flange [1]. One tool is the wheel drive unit; the other is the soil preparation tool, and swapping between them is what allows a full test matrix to run unattended.

Force control is the design idea. Each Cartesian axis is either force controlled or holonomically constrained, so a run holds a commanded vertical load, a commanded slip ratio and a commanded side-slip and tilt while the wheel traverses the bin [1][3]. That removes the hanging weights and cable winch of the conventional bench and replaces them with a commanded set point, which is why arbitrary combinations of load and slip can be swept automatically [1][4]. Constant load and forced slip are held through the robot’s 4 ms control cycle [2].

The published limits do not fully agree. Wheel radius is 80 to 250 mm in the facility paper [1], while the drive unit is separately described as accepting wheels of 200 to 400 mm diameter [2]; both are recorded here. The servo drive is limited to 28 Nm and 85 rpm, giving 1.1 m/s at a 125 mm radius, and the operator page states a wheel load to 1 kN, a run length up to 3 m and a horizontal speed above 1 m/s [5]. The rig was designed to a 6500 x 6500 x 3500 mm envelope with a 10 kN/m2 floor load allowance for the bins [4]. The lowest loads actually run in the published campaigns are 50 N, chosen “to limit the sinkage over the entire slip range”, and 100 N, described as “representative of a medium-class 6-wheeled rover (mass ~350 kg) under lunar gravity”; the design requirement list goes lower, to 37 N for the LRU rover wheel at lunar gravity.

Soil preparation tool

Parameter	Value
Working volume	Full lane; loosening to 250 mm, mixing to about 200 mm [1]
Simulant or terrain	Recipe differs per simulant: 3 min sand, about 9 min chalk
Instrumentation	Bulk density sampler, plus or minus 2.5 percent relative [3]

The second tool on the changer prepares the soil, which is why the test matrix can run without an operator in the bin. Preparation is automated in three steps: loosening, leveling and compression [1]. For mechanical loosening the tool is inserted to a depth, tilted to an angle, traversed, reversed to the opposite angle and returned, which counts as one pass; the published recipe for the cohesive simulant is 20 degrees, 250 mm depth, 0.2 m/s, three passes. Pneumatic loosening by fluidization through a floor grid takes 1 to 10 s per pulse, with 5 s stated as good for olivine and quartz sand, but the operators record two disadvantages, “the dust emission and a possible separation of the different grain sizes”, the separation developing “on a time scale of minutes” [1]. Leveling is a 30 degree blade pass at a set height, and compression is repeated vertical presses of 5 to 20 mm for the olivine and 10 to 80 mm for the chalk, with 50 percent overlap described as “a good balance between effort and quality” [1].

Preparation time is asymmetric between the two simulants: about 3 minutes for the olivine sand, using fluidization and two leveling passes, against almost three times longer for the cohesive chalk, which needs mixing, at least four leveling passes at decreasing depths and two compression passes [2].

Relative density is controlled through measured bulk density rather than through a procedure alone. The method uses a purpose-built non-disturbing sampler with a stated plus or minus 2.5 percent relative measurement uncertainty, sampled along the lane before and after preparation [3]. On the conventional DLR bench, from which the method was carried over, the four preparation methods gave densities of 1.11 g/cm3 by manual deep raking, 1.26 by manual board-walk compression, 1.13 by semi-automated plow and 1.26 by semi-automated sled, with root mean square deviations of 1.6 percent, 2.3 percent, 1 percent and 0.5 percent respectively.

Soil bins

Parameter	Value
Working volume	Small front bin 4 m long, about 3 m usable; large fixed bin behind [1]
Simulant or terrain	RMCS13, 1 to 30 um chalk; RMCS14, 90 to 700 um olivine [2][3]
Instrumentation	Penetrometer, 10 points, 0.2 m depth at 0.02 m/s [2]

There are two containers, a large fixed one behind the robot and a smaller movable one in front, and only the movable one is dimensioned in the literature: 4 m long, taking wheels up to 200 mm wide, of which “due to boundary effects on both sides, the resulting test length is of about 3 m” [1]. Bin depth is not published for either; mechanical loosening reaches 250 mm and sand mixing about 200 mm, so the working depth is at least 0.25 m [1]. Neither tonnage is published.

DLR-RMCS13 is a calcium carbonate powder of 1 to 30 um with rounded ellipsoidal grains that form breakable clusters, cohesive and highly compactable, designed as a worst-case soft soil after Spirit’s embedding: uncompacted it “cannot support the weight of an average person, resulting in knee-deep sinkage”, while compacted “it is even possible to walk on the material without noticeable sinkage” [3]. DLR-RMCS14 is an olivine sand of 90 to 700 um with a friction angle of about 35 degrees, a quartz substitute with no free silica, matched to martian grain size distributions less the silt and clay fractions [2].

Instrumentation

The force-torque sensor does three jobs at once: it measures the wheel forces, it protects the rig against limit violations, and it closes the force control loop [1]. Its IP65 and IP68 rating is a facility choice rather than a sensor choice, because the sensor lives in the dust.

Sinkage is not measured by a displacement transducer on the wheel carrier. It is derived from robot position combined with laser profile scans of the surface taken before and after each run, which the rig reconstructs into a full three-dimensional surface [1]. That gives sinkage, rut geometry and soil displacement from the same measurement. Beyond the force-torque sensor and the two Keyence LJ-7300 scanners, the rig carries an inertial measurement unit, camera mounts, environmental sensors for temperature and for air and soil humidity, and mounting points on the experiment tool for a bevameter and a penetrometer. A Sony NEX-FS700 records 4K at 480 frames per second, and the hall lighting exists to support it rather than to set an illumination level [1].

There is no motion capture volume. Pose comes from the robot’s own telemetry over the KUKA robot sensor interface at a 4 ms cycle [2]. Sampling of the force channels was designed around grouser excitation, at 50 Hz or better against 10 Hz on the previous rig [4].

Soil state is instrumented separately from the wheel. Penetrometer tests are “frequently conducted in ten different spots equally distant from each other to verify that a consistent soil state was achieved”, to 0.2 m depth at 0.02 m/s, plotted against bin position, with transient regions at both ends shaded and their data discarded [2]. An internal consistency check on the 2019 campaign compared a computed grouser strike frequency of 1.193 Hz against 1.19 Hz measured by fast Fourier transform [1].

What it does not reproduce

Gravity. The rig scales the wheel load, not the gravitational field. The 100 N case is described as representative of a 350 kg six-wheeled rover under lunar gravity [2], but the soil column, the wheel’s own weight distribution and the granular flow are all at 1 g. For the milli-gravity case DLR’s published position is blunt: “due to large differences between the Phobos and Earth environments, it is almost impossible to perform laboratory experiments to design, develop and validate robotic operations. Thus, simulations are critical” [6].

Whether a 1 g rig can stand in for that difference at all, even by scaling the load, is unsettled elsewhere in the literature [10]. A parabolic-flight comparison that put the same wheel and simulant through both an equal-load 1 g test and true lunar gravity found the 1 g test overstated mobility, drawbar pull to weight fell 20 percent or more and sinkage rose up to 40 percent in true lunar gravity, and traced the gap to a fourfold drop in measured cone index gradient, a soil shear-strength effect that lowering the density of a 1 g simulant does not reproduce [10]. A discrete element study of a lunar wheel-soil model reports the same direction of effect: simulated lunar gravity gave a less visible wheel track and a higher tractive efficiency than the same wheel and soil model run at 1 g [11]. TROLL’s published campaigns do not apply either correction; the 100 N case stands for lunar gravity by matching load alone.

A quarter of its own bin. The soil bin is 4 m long and the usable test length is about 3 m, because of boundary effects at both ends [1]. The 3 m driving distance in the high-speed campaign was chosen “to avoid potential disturbances caused by the front and back walls of the soil bin”, and only the 0.5 to 2.5 m interval is treated as steady state [2].

A disturbance-free carrier. The arm is part of the measurement. The high-speed campaign records “slightly higher vibrations experienced at both ends of the soil bin where the robot arm was fully extended” [2], which is where the arm’s stiffness is lowest and where the end-effects of the bin already apply.

Its own speed and slip limits. The operators are specific about what fails first: “the main controller of the robot arm had to constantly compensate for the rotational speed of the wheel, which proved to be increasingly difficult when higher speeds or higher slip ratios were considered”, and “greater sinkages may have been experienced due to slightly over-imposed slipping conditions resulting from the controller throughput being challenged to keep pace with the motor driver”, alongside “the close approach in several instances to the motor current limit” [2]. The limits bite above about 1 m/s and above 80 percent slip.

Fluidized preparation on cohesive soil. On the cohesive chalk simulant, fluidization “tended to form cylindrical air channels, releasing the air directly outside of the bin and precluding its use” [2], so the fast preparation route is available only for the sand.

The densest achievable soil state, by machine. “The maximum density of the material was not possible to be reached via semi-automated preparation, but by its manual counterpart” [3]. The automated method that makes the throughput possible does not reach the state the manual method reaches.

Homogeneity along the lane. The operators record that “maintaining the soil in a homogeneous state at any given point was crucial and often a challenge” [2], which is why the ten-point penetrometer check exists.

Traction on the cohesive simulant above 0.1 m/s. The high-speed campaign reports complete loss of traction on DLR-RMCS13 above that speed [2].

Vacuum, temperature and illumination. None of the three is provided. The enclosure carries sensors for temperature and for air and soil humidity, used to keep ambient conditions controlled rather than to set them; logged values for the 2019 campaign were 26.5 C and 1.9 percent relative humidity on one simulant and 24.5 C and 3.3 percent on the other [2].

How TROLL compares to other single wheel testbeds

The gap between TROLL’s rigid and flexible wheel results widens the same way the design space elsewhere predicts. Tohoku University’s rutting work on a rail-carriage testbed finds that both rut depth and rut incline grow with vertical load [7], and its cohesion comparison finds that the DLR-RMCS13-style cohesive simulant and a sand simulant give systematically different wheel performance under the same testbed [8], which is the same pair of variables, load and cohesion, that TROLL’s own campaigns vary through the soil preparation tool rather than through the load path. Carnegie Mellon’s shear-interface imaging attributes a grousered wheel’s gain over a smooth one to reduced bulldozing resistance at the front of the contact patch rather than added thrust at the rear [9], a distinction TROLL’s force-torque instrumentation records as a net drawbar pull number without separating the two contributions.

Two of the open questions the facility page cannot itself answer already have partial answers elsewhere in the literature. A discrete element model of a lunar wheel-soil system, run at gravities from 1 g to 6 g, finds that raising gravity deepens the wheel track and lowers tractive efficiency [11], which is the mechanism a 1 g test at TROLL cannot access no matter how the load is set. A parabolic-flight thesis that measured the same wheel and soil pair at 1 g and at lunar gravity develops granular scaling laws as a correction for exactly that gap, reporting errors of a few percent on drawbar pull and sinkage once wheel size and speed are rescaled between gravity levels [10], a correction TROLL’s own campaigns do not apply. The flexible wheel geometry both testbeds share traces back to the original ExoMars suspension trade study, whose soil-contact model is the reference against which the wheel’s later single-wheel and breadboard performance was checked [12].

Campaigns run there

Side-slip in cohesive soil, 2019. A three-dimensionally printed wheel of 160 mm radius and 100 mm width with twelve straight grousers of 10 by 5 mm, on DLR-RMCS13 prepared without compression, at 0.1 m/s and 50 N, sweeping slip from 20 to 80 percent and side-slip angle across 0, 5, 10 and 15 degrees, with three repeats of each condition [1]. The campaign ran 115 tests in three days with one operator, against several weeks for the same matrix on a conventional manual bench, and three identical runs agreed in both traction and sinkage [1].

High speed tractive performance, run 2019 and published 2023. A rigid metallic wheel of 250 mm diameter and 112 mm width with twelve toothed grousers, and the flexible ExoMars and Rosalind Franklin wheel of the same size and grouser design, at 100 N across 0.01 to 1 m/s at 20 and 60 percent slip, over 3 m after a 0.75 s acceleration, on both simulants [2]. Drawbar pull and tractive efficiency fall sharply above 0.2 m/s; peak efficiency drops by 62.5 percent between 0.01 and 1 m/s; the slip at which efficiency peaks migrates from 10 percent to 30 percent; and sinkage rises with speed, against the prior literature. The flexible wheel at 1 m/s gives about twice the drawbar pull and efficiency of the rigid one with 18 percent less sinkage at low slip [2]. That flexible wheel descends from the 3-lobed titanium and steel design the original ExoMars trade study chose over four competing suspension concepts; the same trade study’s soil-contact model, validated against single-wheel tests, carried 20 to 40 percent drawbar pull error for flexible wheels against 20 percent or less for rigid ones [12], the same order of uncertainty the TROLL facility paper admits for its own servo drive figures. The recommended operating band is 0.2 to 0.3 m/s. The campaign happened at TROLL because no conventional rig could reach those speeds: researchers from Tohoku University and ESA “approached the Institute of System Dynamics and Control” for exactly that reason, which is also the only documented external access route [2].

MMX rover. The operators state that “in the future the testbed will be used within the MMX Rover mission context to test, validate and optimize the rovers wheels” [1]. No report of an executed MMX wheel campaign at TROLL is published; the available MMX locomotion paper covers simulation and the milli-gravity technology experiment rather than TROLL runs [6].