A motor feeding a load through a gear train with measurable backlash behaves like a piecewise system: torque is transmitted only when the motor and load angles differ by more than the backlash half-width h, otherwise the teeth are out of contact and no torque passes — the formal condition stated in MIT course 2.017J, problem 32 [S2].
For that class of mechanism, the choice between a standard AC motor and a servo motor is decided by four hard numbers: closed-loop position error in rad, peak torque in Nm, the gear half-width h, and the tooth-contact stiffness k at the mesh — not by the motor's nameplate alone [S2].
Backlash Mechanics and the MIT Reference Model
Backlash is the decoupling of gear teeth that comes from imperfect meshing, modelled as a dead-zone of half-width h on each side of contact; torque is zero when |θm − θl| ≤ h, and rises linearly with stiffness k once the gap closes, so τt = k(θm − θl − h) for θm − θl > h [S2]. The same problem set gives a concrete working point: Jm = 0.003 kg·m², Jl = 0.01 kg·m², k = 10,000 Nm/rad, h = 0.02 rad, and feedback gain g = 6 Nm/rad, with load damping b treated as a uniform random variable in [0, 0.02] Nm/(rad/s) [S2].
Across that uniformly-sampled range of b, the published run reports a mean load-side position error of 0.0101 rad (≈0.58°) with a standard deviation of 0.00527 rad, and a mean commanded motor torque of 0.0609 Nm (σ = 0.0316 Nm) — values that define the floor a real drive must beat on a backlash-loaded axis [S2]. Any servo drive that cannot hold the load within roughly 0.005 rad of command under the same k, h and g will underperform a correctly tuned induction drive on the same gearbox.
Why an AC Induction Motor Struggles on the Same Gearbox
A standard AC motor without a position feedback loop has no way to measure θl, so the controller in the MIT problem — τm = −g·θl — cannot be closed; the motor simply runs at its V/Hz curve and the gear's backlash dead-band translates directly into load-side drift and reversal spikes. Induction machines also have lower torque density per ampere than permanent-magnet servos, so achieving the 0.0609 Nm mean demanded by the reference run requires either a physically larger frame or a gearbox with a lower reduction ratio, which changes k and h in the model. [S2]
The MIT run quantifies the second cost of staying open-loop on a backlash train: the standard deviation of motor torque is 52% of the mean (0.0316 / 0.0609), so an uncompensated AC drive spends significant frame size on torque ripple rather than on useful work [S2]. A correctly specified servo system with current-loop bandwidth above the natural frequency √(k/Jl) = √(10,000/0.01) ≈ 1000 rad/s (~159 Hz) collapses that ripple, but only if the drive, encoder and gear mesh are sized as a chain [S2].
Where the Servo Motor Wins on Repeatability
Servo motors specified for backlash-loaded axes come with incremental encoders typically at 20-bit resolution (1,048,576 counts/rev, ≈0.0006°/count), absolute multi-turn encoders at 23-bit, and current-loop bandwidths in the 1–3 kHz range — values that let the controller resolve the dead-zone crossing before the load overshoots. Commercial lines such as the FANUC αi-D, αi-B, βi-B, the DD D i S-B direct-drive series, and the L i S-B linear series are explicitly engineered to be paired with CNC position loops, with the motor lineup described as covering small to large machines and sized jointly with the CNC to optimise machine-tool performance [S3].
For backlash-sensitive axes, the servo is normally paired with a low-backlash planetary gear coupling stage rather than a standard spur reducer: planetary heads advertised for servomotors and steppers use a multi-stage ratio and high-efficiency output torque to push the mesh into the arc-minute class, while still accepting the servo motor's shaft envelope [S5]. When the application can tolerate direct drive, the DD D i S-B series removes the gear entirely and pushes the dead-band down to the encoder resolution, which is the cleanest fix for the MIT model's h = 0.02 rad problem [S3].
Decision Matrix: Servo vs AC on a Backlash Axis
Use a servo motor when repeat position error must stay below roughly 1 arc-min (≈0.00029 rad), peak torque is below the servo frame's continuous rating, the axis is reversing frequently (so the dead-zone is crossed many times per minute), and the gearbox can be specified as low-backlash or replaced by direct drive [S2][S3][S5]. Use a standard AC induction motor with a VFD when the application is constant-speed (fans, pumps, conveyors), reversal rate is low, and the gear train's backlash does not feed back into a positioning tolerance — the VFD handles the speed envelope but cannot close the position loop inside the MIT dead-band [S6].
Cost-wise, an AC motor plus a general-purpose VFD is materially cheaper per kW than a matched servo motor, drive, encoder cable and low-backlash gearbox, so the crossover favours servo only when the value of tighter positioning exceeds the hardware delta. Servos also need commissioning (gain tuning at the k/Jl natural frequency) and qualified spares, topics that interact with broader robotic-cell architecture choices such as those covered in Industrial Ethernet Switch Selection for Robotic Workcells: Five Hard Criteria and bearing selection in Thrust Bearing Selection Criteria for Servo Positioning Axes [S3].
Limits, Failure Modes and Sourcing Standards
Three failure modes dominate backlash-loaded axes: (1) limit-cycle chattering at roughly 2h peak-to-peak when the loop gain g is too high relative to the tooth stiffness k; (2) lost motion on direction reversal equal to 2h before the opposite tooth flank makes contact — a direct consequence of the model's dead-zone; and (3) torque-ripple-induced heating, which the MIT run quantifies as a 52% σ/mean ratio on commanded torque when b is uncertain [S2]. The chattering limit is bounded by choosing g below the critical gain 4·k·Jl·h, a stability margin that any integrator-based servo drive should preserve.
Specifying the hardware means locking four numbers: encoder resolution in bits or arc-sec, current-loop bandwidth in kHz, gear backlash in arc-min, and tooth-contact stiffness k in Nm/rad. Gear-head data sheets for servomotor duty should state backlash in arc-min, efficiency above 0.90 per stage, and rated torque at the output shaft, with a multi-stage planetary ratio where the first stage handles the servo-side input [S5]. For OEM packages like the FANUC αi-D, the matching CNC handles the loop closure, but the gear-head outside the motor is still the specifier's responsibility [S3]. Use the MIT 2.017J reference run as a sanity check: any candidate system that cannot hold load-side position error within roughly 0.01 rad under k = 10,000 Nm/rad and h = 0.02 rad is not fit for a backlash-loaded positioning axis [S2].
Selection Checklist and Trackable Signals
Confirm in writing: backlash half-width h ≤ 0.005 rad (≈17 arc-min) for general positioning, ≤ 0.001 rad for high-precision cells; tooth-contact stiffness k published on the gear-head data sheet; servo drive current-loop bandwidth ≥ 1 kHz; encoder ≥ 20-bit incremental or 23-bit absolute multi-turn; gearbox efficiency ≥ 0.90 per stage [S2][S3][S5]. Reject any gear-head offered for servo duty without a published arc-min backlash figure, and reject any AC drive offered for a reversing positioning axis without an external position feedback device. Track the next two signals: a vendor's published bandwidth on a low-backlash planetary head at 1-stage and 2-stage ratios, and the rolling release of matched CNC + servo bundles that size the motor and drive together rather than as separate line items [S3][S5].