Robotics Platforms & Direct Drive Motors | Direct Drive Tech

A direct-drive mobile robot removes the gearbox between each traction motor and wheel, so the motor must produce full wheel torque at operating speeds that often remain below 200 rpm. For a 300 kg robot with 160 mm wheel radius, 1 m/s² acceleration and a 5% grade, two motors may each need about 40 N·m before safety allowance. The design therefore moves toward wider multipole motors, higher phase current, stronger wheel bearings, high-resolution encoders and better chassis cooling. Gear backlash, lubrication and gear wear disappear, but winding heat, regenerative braking, zero-speed control and structural alignment need more engineering work.

A conventional mobile robot often uses a compact motor running between 2,000 and 6,000 rpm, followed by a planetary, spur or helical gearbox. A direct-drive layout removes that reduction stage and connects the motor rotor to the wheel hub, either through a short shaft or through an integrated outer-rotor structure.

Removing one assembly changes more than the number of mechanical parts. A 20:1 gearbox allows a motor producing 2 N·m to deliver close to 40 N·m before transmission losses, while a gearless motor must produce the full 40 N·m itself.

That torque requirement changes motor diameter, winding selection and current capacity. It also changes how the chassis supports wheel forces because the motor housing can no longer be treated as a lightly loaded component mounted behind a separate wheel bearing.

The architectural differences can be seen in a simple comparison:

Design area Geared traction unit Direct-drive traction unit
Typical motor speed 2,000–6,000 rpm 50–300 rpm
Reduction ratio 10:1–40:1 1:1
Motor torque for 40 N·m wheel torque 1–4 N·m About 40 N·m
Mechanical backlash Often 0.2°–1.5° at the output Near zero from transmission parts
Transmission efficiency Commonly 75%–95% per operating point No gearbox loss
Wheel-position sensing Motor encoder multiplied by ratio Encoder reads wheel rotation
Holding when unpowered May receive resistance from gearing Usually back-drivable
Routine transmission service Lubrication, wear checks or replacement No gear service
Main packaging pattern Long and narrow Wide and short

A design should begin with wheel force rather than motor catalog power. Motor wattage alone says little about whether the robot can start on a slope, rotate on a high-friction floor or cross a 15 mm threshold at full payload.

Consider a 300 kg differential-drive robot using two powered wheels with a radius of 0.16 m. Assume 1 m/s² acceleration, a 5% grade and rolling resistance equal to 2% of vehicle weight.

The acceleration force is 300 N. Grade force is about 147 N, and rolling resistance contributes about 59 N, giving 506 N before turning resistance or surface impact is added.

Dividing the force between two wheels gives 253 N per wheel. Multiplying by the 0.16 m radius produces about 40.5 N·m per wheel.

A 25% engineering allowance raises the short-duration requirement to roughly 51 N·m. A larger 50% allowance raises it to about 61 N·m, which may be appropriate when floor joints, tire deformation and payload placement are not tightly controlled.

A motor rated for 60 N·m peak and 18 N·m continuous may pass a short acceleration test but overheat during a 10-minute grade climb. Peak torque and continuous torque must therefore be checked as separate operating limits.

The same robot travelling at 2 m/s turns each 0.16 m radius wheel at about 119 rpm. A geared motor may run near 2,380 rpm with a 20:1 reduction stage, while the gearless motor remains at 119 rpm.

Mechanical power at 40.5 N·m and 119 rpm is about 505 W per wheel. The figure appears moderate, but the motor must produce it at a speed where cooling airflow is low and copper loss may dominate.

Low rotational speed explains why direct-drive motors are usually wider than ordinary high-speed motors. Torque rises with tangential force and effective air-gap radius, so increasing motor diameter allows the electromagnetic force to act farther from the axis.

A wider rotor can produce more torque without increasing axial length by the same percentage. This produces the flat, ring-shaped or “pancake” geometry seen in many frameless torque motors and integrated wheel motors.

A 70 mm high-speed motor may fit behind a gearbox, but a gearless wheel for the same robot could require a motor diameter of 180–300 mm. Exact dimensions depend on magnetic loading, cooling, winding temperature and allowable current density.

Increasing diameter introduces packaging limits. A robot using 150 mm wheels cannot place a 200 mm motor completely inside the wheel, so the motor may need to sit inboard, use a larger wheel or become part of a structural hub extending beyond the tire.

Larger wheels reduce torque for the same tractive force only when radius decreases, not increases. A larger wheel requires more shaft torque for the same ground force, although it may cross obstacles more easily and rotate more slowly.

For example, 250 N of tractive force requires 30 N·m with a 120 mm radius wheel and 45 N·m with a 180 mm radius wheel. The larger wheel raises torque by 50% while reducing wheel speed by 33% at the same vehicle speed.

Motor geometry must therefore be chosen together with tire diameter, floor clearance and threshold requirements. Selecting the wheel first and the motor later can create a package that is either too wide or unable to meet continuous torque.

Outer-rotor motors often suit wheel integration because the rotor shell can attach to the hub while the stator remains fixed to the chassis. The large rotor radius also helps produce low-speed torque.

An inner-rotor motor can still be used, particularly when bearings and wheel shafts must remain conventional. Its smaller electromagnetic radius may require greater axial length, more current or a larger active motor volume.

Axial-flux motors provide another option when wheel width is restricted. Their magnetic flux crosses the air gap along the shaft axis, allowing a thin disc shape with a comparatively large working radius.

Each geometry creates a different thermal path. An outer rotor places rotating magnets near the housing surface, while the heat-producing stator remains fixed and can often be connected to an aluminum carrier.

A frameless motor removes the manufacturer’s housing, shaft and bearing set. That can reduce duplicated metal, but the robot structure must maintain the rotor-to-stator air gap under wheel force, temperature change and assembly tolerance.

A radial air gap of 0.5–1.0 mm leaves little room for bearing movement. A 0.2 mm shaft deflection may consume 20%–40% of the available clearance and increase local magnetic force.

Concentricity should therefore be controlled at the motor mounting surfaces, not only at the tire. Runout introduced by the hub, bearing seat and rotor carrier adds together.

One practical tolerance stack might allocate 0.05 mm to the bearing seat, 0.05 mm to the rotor carrier, 0.05 mm to assembly position and 0.05 mm to operational deflection. The total worst-case movement reaches 0.20 mm before thermal growth is counted.

The wider motor also produces higher overturning moments at the bearing set when the tire contact patch sits outside the bearing span. Moving the wheel center 50 mm farther from the outer bearing increases bending moment by 50% for the same radial force.

Bearing selection therefore moves earlier in the design process. Standard motor bearings are often sized for rotor support, not for a 150 kg wheel reaction combined with curb impact and lateral tire scrub.

A wheel bearing assembly may need to support radial force, axial force and moment at the same time. Paired angular-contact bearings, tapered roller bearings or crossed-roller bearings may be considered when axial stiffness and compact packaging are required.

Bearing catalog capacity is not the same as service life inside a mobile robot. Shock, contamination, preload error and hub deflection can reduce usable life even when calculated steady force remains below the catalog rating.

A 300 kg four-wheel robot may place about 736 N on each wheel when weight is evenly distributed. A payload shifted toward one corner can raise one wheel reaction by 20%–40%, depending on chassis stiffness and center-of-mass position.

Crossing a floor joint can briefly multiply that force. A design check using 2g vertical acceleration raises a 736 N static reaction to about 1,472 N before lateral force or bearing moment is included.

Outdoor platforms may use 3g–5g structural checks because rocks, roots and curb edges produce sharper inputs. The motor rotor, magnets, encoder target and cable exits must survive the same event.

Wheel impact also reaches the stator structure without a flexible gearbox or long shaft to absorb part of the disturbance. Structural stiffness becomes closely connected to motor performance because air-gap change can alter torque ripple and local heating.

Torque ripple deserves attention at low speed. A gearbox may reduce or average some motor pulsation at the wheel, but a gearless rotor passes it to the tire without a reduction stage.

A motor with 3% torque ripple produces a 1.8 N·m variation while delivering 60 N·m. With a 0.15 m wheel radius, that variation equals 12 N of changing ground force.

On a 250 kg robot, 12 N can create about 0.048 m/s² of acceleration variation before tire compliance and vehicle mass filter the response. During a 20 mm/s docking movement, the effect may appear as visible speed steps.

Motor slot count, pole count, magnet shaping and current waveform all influence the result. A lower-cogging motor may use skewed magnets, fractional-slot windings or slotless construction, although each approach changes cost, torque density or manufacturing difficulty.

Current control then handles the remaining ripple. Field-oriented control normally regulates phase current as torque-producing and flux-related components, allowing smoother low-speed operation than simple six-step commutation.

The current loop commonly runs at 10–40 kHz, while the speed loop may run at 500 Hz–2 kHz. A navigation controller may update wheel commands at only 50–200 Hz.

Separating the loop rates helps the motor controller respond to electrical and mechanical changes before the higher-level computer sends a new path command. A 20 kHz current loop executes 100 updates during one 5 ms navigation interval.

Current sensing accuracy affects wheel-force accuracy. A 1% current measurement error becomes roughly a 1% torque estimation error when the motor torque constant is stable.

For a 60 N·m command, 1% equals 0.6 N·m. Two wheels with opposite measurement offsets could differ by 1.2 N·m, causing heading error during long straight travel.

Offset calibration, gain matching and sensor temperature drift therefore matter in a gearless system. Hall-effect current sensors may provide isolation and high range, while shunt resistors can provide better linearity at lower cost if layout and common-mode voltage are handled correctly.

Motor torque is approximately torque constant multiplied by torque-producing current. A motor with a torque constant of 1.5 N·m/A needs about 40 A to produce 60 N·m.

If phase resistance is 80 milliohms, a simplified three-phase copper-loss estimate near that operating point can exceed 190 W, depending on current definition and modulation. Resistance also rises as the winding heats.

Copper resistance increases by about 0.39% per degree Celsius. A winding that rises from 25°C to 105°C experiences an increase of roughly 31%.

An 80 milliohm phase resistance at 25°C can therefore approach 105 milliohms at 105°C. The same current then produces about 31% more copper heat, so temperature can rise faster near the upper operating range.

This relationship makes thermal design part of torque sizing. A motor that appears adequate from a room-temperature torque curve may lose continuous capability inside a sealed robot operating at a 40°C ambient temperature.

The most useful thermal path usually runs from stator laminations and windings into a stationary aluminum carrier. A large contact diameter, thin interface layer and flat mounting surface reduce thermal resistance.

A 0.2 K/W motor-to-chassis path produces a 40°C temperature increase at 200 W of loss. A weaker 0.6 K/W path produces a 120°C increase from the same heat.

Potting material can improve conduction from winding end turns, but it may also make repair impossible and introduce thermal-expansion stress. Its conductivity, cure shrinkage, glass-transition temperature and adhesion should be reviewed rather than choosing material by hardness alone.

Air cooling helps less at low wheel speed because the motor cannot rely on a shaft-mounted fan. A separate blower may use 10–40 W and add filters, ducts and acoustic noise.

Liquid cooling can support higher continuous torque but adds pumps, hoses and leak management. It is more common on heavy outdoor robots or platforms where each wheel must provide several kilowatts for long periods.

Temperature sensing should include at least one winding sensor and one inverter sensor. A housing sensor alone may respond 30–120 seconds later than the winding during a short high-current event.

A thermal model can estimate winding temperature between sensor updates. The model may use measured current, estimated resistance, motor speed, ambient temperature and chassis temperature.

A useful validation program can include 30 repeated acceleration cycles, a 20-minute grade run and a 60-minute stop-and-go route. Temperature should reach a repeatable plateau or remain below the insulation and magnet limits with the planned allowance.

A motor that survives one 10-second peak event has not demonstrated a usable duty rating. Heat accumulation over 20 or 50 repeated events gives a more useful picture of fleet operation.

Electrical sizing extends beyond the motor. Four wheels each requesting 40 A can place a 160 A demand on the battery bus, excluding steering, computing, sensing and payload equipment.

At 48 V, 160 A equals 7.68 kW of electrical input. A 5% voltage drop across cables, contactors and connectors removes about 384 W before energy reaches the inverters.

Cable resistance must therefore be kept low without making wheel modules difficult to service. A removable module may use one high-current connector, one communication connector and one low-voltage safety connection.

Connector ratings should be checked at actual enclosure temperature. A connector rated for 60 A in 25°C open air may require derating inside a 55°C sealed compartment.

Cable bending also matters when a powered wheel is mounted on suspension or steering. A cable designed for fixed installation may fail after 100,000 bending cycles even though its current rating is adequate.

Flexible conductors should be routed above their minimum bend radius and protected from tire spray, sharp edges and repeated torsion. A steering wheel turning ±120° creates a different cable requirement from a fixed differential-drive wheel.

The battery-management system must support peak discharge and regenerative charge. A battery capable of 200 A discharge may accept only 40–80 A of charge, especially near full state of charge or at low temperature.

Regenerative braking can therefore raise bus voltage faster than the battery can absorb energy. A robot decelerating 300 kg from 2 m/s stores about 600 J of translational kinetic energy before wheel and rotor rotation are counted.

Stopping in 1 second returns an average mechanical power of about 600 W. Stopping in 0.2 seconds raises the average to about 3 kW, with a higher instantaneous peak.

A full battery may reject part of that energy. A braking resistor, active clamp or reduced regeneration command can prevent overvoltage when the battery cannot accept charge.

The resistor must be sized for energy and temperature, not only peak wattage. Ten 600 J stops within 2 minutes place 6,000 J into the braking system if none of the energy returns to the battery.

Braking has a mechanical side as well. A gearless wheel normally rolls when motor current disappears, so a slope-holding method is needed whenever unpowered movement is unacceptable.

Electrical holding torque consumes current and produces heat. Holding 20 N·m with a 1.5 N·m/A torque constant requires about 13.3 A per motor.

Two motors holding for 10 minutes can consume noticeable battery energy while adding winding heat at zero speed. Zero-speed operation also provides no rotational airflow.

A spring-applied, electrically released brake can hold the wheel without continuous power. Its static rating should exceed the maximum slope torque with a stated allowance, often 25%–100% depending on safety requirements and uncertainty.

A brake rated for static holding may not tolerate repeated emergency stops. Friction material temperature, allowable energy per stop and wear life must be checked separately.

Stopping distance also depends on tire friction and floor condition. A robot with 0.5 available tire-floor friction can theoretically decelerate near 4.9 m/s² before slip, but payload stability and passenger-area safety may require a much lower value.

A 2 m/s robot decelerating at 1 m/s² needs about 2 m of theoretical braking distance before control delay is added. At 2 m/s², theoretical distance falls to about 1 m.

A 100 ms control and communication delay adds 0.2 m of travel at 2 m/s. Safety scanning distance must account for that delay, brake build-up time and floor variation.

Encoder selection influences both braking and low-speed travel. A geared motor encoder benefits from the reduction ratio because one wheel revolution produces many motor revolutions.

A 4,096-count motor encoder behind a 20:1 gearbox provides 81,920 counts per wheel revolution before backlash and interpolation are considered. The same encoder mounted on a gearless wheel provides only 4,096 counts.

With a 0.942 m wheel circumference, 4,096 counts equal about 0.23 mm of theoretical travel per count. At 81,920 counts, the value falls to roughly 0.0115 mm.

The lower direct-drive count may still be adequate for navigation, but speed measurement becomes less smooth at very low velocity. At 10 mm/s, the 4,096-count encoder produces only about 43 counts per second.

A 16-bit absolute encoder provides 65,536 positions per revolution, raising the same 10 mm/s signal to about 696 counts per second. The controller can then estimate speed over shorter time windows.

High resolution does not guarantee high accuracy. Eccentricity, magnetic target error, interpolation error and mounting runout can create periodic position error.

A sensor advertising 18-bit resolution may have absolute accuracy of only ±0.1° or ±0.2°. That may still suit wheel-speed control, but the specification should not be treated as 262,144 perfectly accurate positions.

Absolute encoders can report rotor angle immediately after startup. Incremental encoders need an index search or a separate commutation sensor unless the controller can estimate initial rotor position.

For smooth torque at zero speed, the controller needs electrical rotor angle with enough accuracy to align stator current. A 5° electrical-angle error reduces useful torque and creates unwanted current.

The relationship between mechanical and electrical angle depends on pole pairs. A motor with 20 pole pairs turns 1° mechanically while electrical angle moves 20°.

A mechanical sensor error of 0.25° therefore becomes 5° electrically in that motor. Higher pole counts improve low-speed torque geometry but place tighter requirements on position sensing and alignment.

Sensor bandwidth also matters during wheel impact. A measurement filtered too heavily may hide vibration but delay torque control, while insufficient filtering can pass noise into the current command.

Encoder cables should be separated from phase conductors or use differential signaling. Large phase-current changes can couple noise into single-ended feedback lines.

A 40 A edge switching in less than 1 microsecond can create substantial electromagnetic interference. Shield termination, grounding and inverter layout should be planned before the first prototype.

The Direct Drive Tech approach is most useful when motor, encoder, inverter, bearing and chassis interfaces are treated as one wheel assembly rather than as five unrelated purchases.

That integration can reduce part count, but it also concentrates failure effects. A damaged encoder or bearing may require replacement of the complete wheel module if the design does not allow field service.

Modular replacement can keep service time below 30 minutes when connectors, fasteners and software identification are standardized. A built-in serial number and calibration record can allow a new module to load its motor parameters automatically.

Without modular planning, wheel replacement may require rotor alignment, encoder calibration, brake adjustment and cable routing. A two-hour workshop task is unsuitable for fleets expected to operate 20 hours per day.

Gearless systems remove gear lubrication and tooth wear, although tires, bearings, seals, brakes and connectors remain service items. Maintenance work changes rather than disappearing.

Gearboxes may develop increasing backlash over thousands of hours. Direct coupling avoids that wear mechanism, so docking accuracy is less likely to change because of gear clearance.

Bearing wear can still create wheel play. An increase from 0.05 mm to 0.30 mm radial movement may disturb the air gap, encoder alignment and seal contact before the tire appears loose to an operator.

Condition monitoring can compare current, speed and temperature across wheels. Four nominally identical motors operating on a level floor should show similar current after accounting for steering and weight distribution.

A 15% current increase on one wheel over several weeks may indicate tire pressure loss, bearing drag, brake contact or alignment change. The alert threshold should be established from fleet data rather than from one prototype.

A useful baseline can be built from 20–50 robots across at least 100 operating hours each. A sample of two vehicles does not show enough manufacturing and route variation for reliable service thresholds.

Motor resistance can also be checked during startup. A change of 10% after temperature compensation may indicate connector corrosion, winding damage or measurement error.

Vibration sensors may detect bearing defects, although wheel texture and floor joints create strong background signals. Comparing matched route segments is more useful than applying one vibration limit to every floor.

Ingress protection becomes harder when the motor sits beside the tire. Water, cleaning chemicals, fibers and abrasive dust can reach the hub repeatedly.

An indoor warehouse robot may target IP54, while outdoor or washdown platforms may require IP65, IP67 or higher at the wheel module. The chosen rating should match the complete assembled unit, including rotating seals and connectors.

A static enclosure test does not reproduce pressure changes from temperature cycling. A sealed wheel warmed to 70°C and then exposed to 10°C water can draw moisture through small seal defects as internal air contracts.

A pressure-equalization vent can reduce the pressure difference while blocking liquid water. Its flow capacity and contamination resistance need review because dust or detergent can reduce performance.

Shaft seals add friction, which matters more at low speed. If each seal adds 0.5 N·m and the wheel normally needs 5 N·m on a smooth floor, seal friction raises torque demand by 10%.

Low-friction labyrinths reduce contact loss but may not stop high-pressure spray. A combined labyrinth and contact seal often provides better protection, with additional width and assembly work.

Corrosion also affects the motor air gap. Rust particles from untreated steel can attach to permanent magnets and damage insulation or encoder surfaces.

Stainless hardware, coated laminations and controlled drainage help in wet environments. Material selection should account for salt spray if the robot operates near coastal facilities.

Direct coupling changes traction control because motor torque reaches the wheel with little transmission compliance. Current can therefore provide a fast estimate of wheel force after motor losses are characterized.

A 1.5 N·m/A motor receiving 20 A produces about 30 N·m before cogging, bearing and tire effects. With a 0.15 m wheel, that corresponds to roughly 200 N of tangential force.

If the wheel accelerates rapidly while inertial sensors show little vehicle acceleration, slip is likely. The controller can reduce current within 5–20 ms rather than waiting for a large navigation error.

Traction control is useful on polished concrete, wet floors and loading-dock plates. Available friction can vary by more than 50% between clean dry concrete and a contaminated surface.

Torque allocation should also account for weight transfer. During forward acceleration, normal force shifts toward the rear axle, reducing the front wheels’ usable traction.

For a 300 kg robot with a 0.5 m center-of-mass height, 1 m wheelbase and 1 m/s² acceleration, longitudinal weight transfer is about 150 N. That equals roughly 5% of total vehicle weight.

A four-wheel system commanding equal torque may therefore make the lighter axle slip first. Current limits based on estimated wheel normal force can improve stability.

Turning creates another difference. Differential-drive and skid-steer robots force tires to scrub laterally, especially when wheelbase and track width are similar.

Turning resistance can exceed rolling resistance by several times on high-friction flooring. A robot that needs 8 N·m per wheel in straight travel may need 20–35 N·m during a stationary turn.

The exact value depends on tire compound, contact patch, payload position and chassis compliance. Bench motor sizing should therefore include full-payload turning tests on at least three representative floors.

A 30-turn test on smooth epoxy, textured concrete and rubber flooring can reveal whether current limits or thermal limits are reached. One polished laboratory surface cannot represent a mixed warehouse route.

Omnidirectional wheels reduce lateral scrub but introduce roller losses and vibration. Mecanum or omni wheels may need higher motor speed accuracy because small wheel-speed differences create sideways drift.

A 2% speed mismatch between left and right wheels can create a growing heading error. Direct wheel sensing removes gearbox backlash from the control path, but tire diameter tolerance remains.

A nominal 300 mm tire with ±1% diameter variation can differ by 3 mm. Two wheels at the same rpm then travel different distances.

Wheel-radius calibration can correct much of that difference. The robot can compare encoder travel with an external reference over a 10–20 m path and store a scale value for each wheel.

Tire wear changes the calibration over time. A 2 mm reduction in radius on a 150 mm wheel changes travel per revolution by about 1.3%.

Fleet software may therefore recalibrate after tire replacement or after a set distance such as 1,000 km. The suitable interval depends on tire material and floor abrasiveness.

Energy use must be evaluated across a complete route. Removing a gearbox removes gear-mesh, seal and oil-churning losses, but a large low-speed motor may have higher copper loss than a smaller geared motor.

Suppose a gearbox operates at 90% efficiency while delivering 500 W mechanical output. About 56 W is lost in the gearbox.

A direct-drive motor removes that 56 W, but it provides an energy benefit only if its extra copper and inverter losses remain below the removed transmission loss. A poorly matched gearless motor could lose 100–200 W at the same wheel point.

Route measurements should include acceleration, cruise, turning, waiting, docking and slope operation. A robot rarely stays at one rated point for an entire shift.

One representative route might contain 30% cruise, 20% acceleration and braking, 25% turning, 20% waiting and 5% ramp travel. Changing only the ramp share from 5% to 15% can alter motor temperature and battery use noticeably.

Regeneration can recover part of braking energy, although battery acceptance and low-speed control limit the amount. A system recovering 50% of 600 J during each stop returns 300 J.

Across 1,000 stops, that equals 300 kJ, or about 83 Wh. On a robot with a 2 kWh battery, the recovered amount equals about 4.2% of battery capacity.

Real savings depend on route length, rolling loss and auxiliary power. Computers, lidars, cooling fans and payload equipment may consume 100–500 W regardless of traction architecture.

A 10% traction-efficiency improvement does not create a 10% runtime improvement when traction represents only half of total energy use. If traction uses 50% of battery energy, a 10% traction reduction improves total energy use by about 5%.

Noise performance often improves because gear-mesh tones disappear. The result is useful in hospitals, hotels, offices and laboratories where a 5–10 dB reduction can be noticeable.

The motor can still produce tonal noise from cogging, current harmonics and structural resonance. A quiet gearbox removal does not correct an inverter switching tone or a flexible wheel cover.

Switching frequency above 20 kHz moves the main electrical tone beyond much of adult hearing, but switching losses rise. Motor inductance and controller hardware determine whether the higher frequency is practical.

Structural modes should be checked with the motor operating through its full speed range. A 24-pole motor can create force harmonics that excite a wheel cover or chassis plate at specific speeds.

Adding stiffness may work better than adding damping mass. A plate thickness increase from 2 mm to 3 mm can raise bending stiffness by more than 200% because stiffness scales strongly with thickness.

Direct-drive packaging may free central chassis space by moving motors into the wheels. The space can be used for larger batteries, lower payload platforms or wider service access.

Moving motor mass outward also increases yaw inertia. Four 6 kg wheel motors placed 0.45 m from the center contribute about 4.86 kg·m² of yaw inertia before the rest of the vehicle is counted.

The same 24 kg placed 0.20 m from the center contributes only 0.96 kg·m². Outboard motors can therefore make rapid rotation require more torque even when total mass is unchanged.

Suspended outdoor robots also gain unsprung mass. A heavier wheel assembly responds less easily to small surface changes and sends greater impact force into the chassis.

An indoor platform on flat flooring may accept a 10–15 kg wheel module. A small outdoor robot may need a much lighter assembly to keep tire contact over uneven ground.

Cost comparison should include more than the motor price. A direct-drive motor may use more copper, magnet material and precision machining, while removing the gearbox, coupling and some assembly work.

A geared module could contain a $180 motor, $240 gearbox, $60 coupling and $100 encoder. A gearless module might use a $520 integrated motor, $160 encoder and larger $120 bearing set.

The first example totals $580 before housing and assembly. The second totals $800, but it may avoid scheduled gearbox replacement and reduce calibration work.

If a gearbox replacement costs $500 in parts and labor every 8,000 operating hours, a fleet running 4,000 hours per year reaches that event about every 2 years. A 100-robot fleet could therefore face $50,000 per replacement cycle.

The gearless option must still account for wheel-bearing and encoder replacement. A sealed integrated module costing $1,200 to replace may be expensive when only a $60 bearing has failed.

Serviceable bearing cartridges and replaceable encoders can reduce that risk. The extra fasteners and seals may add 5%–10% to module cost but lower field repair cost over several years.

The most suitable architecture depends on measurable requirements rather than preference.

Requirement Direct drive is often suitable when Gearing is often suitable when
Wheel speed Mostly below 300 rpm Motor can operate efficiently above 2,000 rpm
Wheel torque Moderate and enough motor diameter is available Torque is high relative to wheel size
Noise Low tonal noise is required Gear noise is acceptable
Backlash Precise reversals and docking matter Small output play is acceptable
Package shape Wide, short module fits Long, narrow module fits
Holding Separate brake is acceptable Transmission resistance helps
Service plan Module replacement is preferred Gearbox maintenance is available
Operating surface Relatively smooth and controlled Severe shock favors separated motor placement
Thermal path Chassis can absorb motor heat Motor has its own cooling arrangement
Production volume Integration cost can be spread across many units Off-the-shelf assemblies reduce early cost

A prototype program should test more than maximum speed. Low-speed crawling, repeated turning, grade holding, emergency braking and thermal recovery often expose problems that a straight-line speed run misses.

A practical program may use 50 acceleration cycles at maximum payload, 30 stationary turns on three floor surfaces, a 60-minute route test and 20 emergency stops at both 50% and 100% battery charge.

Motor winding, inverter, connector and bearing temperatures should be recorded. Current, bus voltage, wheel speed and chassis acceleration should use synchronized timestamps.

Acceptance limits should include continuous temperature, short peak temperature, bus overvoltage, wheel-speed error, stopping distance and current imbalance. A single “test passed” label does not show which margin remains.

Production variation should also be measured. Testing 10 wheel modules can reveal resistance, torque constant, encoder offset and bearing-friction spread.

If motor resistance varies by ±5% and torque constant by ±3%, current-loop and thermal performance will not be identical across the fleet. Calibration or software allowance may be required.

Environmental checks can include temperature cycling, vibration, dust exposure and water spray at the planned enclosure rating. A 2026 design intended for five years of operation should also consider replacement-part availability through 2031.

Supplier documentation should state continuous torque at a defined mounting temperature, peak torque duration, winding temperature limit, magnet temperature limit, encoder accuracy, bearing arrangement and ingress rating.

A torque number without speed, temperature and duration is incomplete. “60 N·m peak” could describe 0.5 seconds, 10 seconds or 60 seconds, with very different use inside a mobile robot.

The final wheel-module specification should list at least the following items:

  • Continuous wheel torque at the highest planned ambient temperature

  • Peak torque and maximum allowed duration

  • Maximum mechanical and electrical speed

  • Phase and bus current limits

  • Winding resistance at 25°C and hot operating temperature

  • Torque constant and back-EMF constant

  • Encoder resolution, accuracy and communication method

  • Bearing radial, axial and moment capacity

  • Brake static torque and permitted stopping energy

  • Motor-to-chassis thermal resistance

  • Maximum winding, magnet, bearing and inverter temperatures

  • Water, dust, chemical and corrosion protection

  • Cable bend life and connector current rating

  • Rotor and hub runout limits

  • Wheel replacement and calibration procedure

  • Measured efficiency at no fewer than 10 torque-speed points

A robot meeting those requirements can gain low backlash, lower gear noise and fewer transmission service items without accepting uncontrolled heat or braking behavior. A robot missing the thermal, bearing or current checks may remove one gearbox and create several harder faults inside the wheel.