Skip to content

Auxiliary IMU factor

Use this page to see how MVIS calibrates every IMU after the first one.

An auxiliary IMU is any IMU after the first in input.imu_topics. It is rigidly mounted, so it feels the same motion as the base IMU, seen from another place, along other axes and on another clock. MVIS solves, for each one:

  • Extrinsic IsjIT{}^{I}_{I_{s_j}}\mathbf{T}: its rotation and position in the base-IMU frame. The position part is its lever arm.
  • Time offset td,jt_{d,j}: an auxiliary sample stamped tt is the motion at base-IMU time t+td,jt + t_{d,j}.
  • Its own intrinsics: Dw\mathbf{D}_w, Da\mathbf{D}_a, Tg\mathbf{T}_g, in the same convention as the base IMU.
  • Its own t_ga and its own biases.

Below we write IsI_s for IsjI_{s_j}. Symbols follow Notation & glossary. The base factor this builds on is Base IMU factor.

VariableSizeWhere
calibration: IsIT{}^{I}_{I_s}\mathbf{T}, td,jt_{d,j}, 24 intrinsics6 + 1 + 24 = 31one per auxiliary IMU
velocity GvIs{}^{G}\mathbf{v}_{I_s}, biases bg,s\mathbf{b}_{g,s}, ba,s\mathbf{b}_{a,s}9at the auxiliary knots
t_ga of this IMU1one per auxiliary IMU

The auxiliary IMU has no pose states of its own. Its poses come from the base trajectory, so every auxiliary factor pulls on the base states too.

Knots: one auxiliary factor spans 3 base intervals by default (solve.aux_imu_stride: 3). Fewer factors make the solve faster. The stride is not accuracy-neutral: the lever arms shift with it. When the lever arm matters, solve again with solve.aux_imu_stride: 1 (one factor per interval, slower) and compare.

At each knot kk MVIS builds the auxiliary IMU’s pose from the base state, the extrinsic and the time offset:

IsGRk=IkGR  Exp ⁣(IsIR  Isωk  δtd)  IsIR{}^{G}_{I_s}\mathbf{R}_k = {}^{G}_{I_k}\mathbf{R}\;\mathrm{Exp}\!\big({}^{I}_{I_s}\mathbf{R}\;{}^{I_s}\boldsymbol{\omega}_k\;\delta t_{d}\big)\;{}^{I}_{I_s}\mathbf{R} GpIs,k=GpIk+GvIk δtd+IkGR  Exp ⁣(IsIR  Isωk  δtd)  IpIs{}^{G}\mathbf{p}_{I_s,k} = {}^{G}\mathbf{p}_{I_k} + {}^{G}\mathbf{v}_{I_k}\,\delta t_{d} + {}^{G}_{I_k}\mathbf{R}\;\mathrm{Exp}\!\big({}^{I}_{I_s}\mathbf{R}\;{}^{I_s}\boldsymbol{\omega}_k\;\delta t_{d}\big)\;{}^{I}\mathbf{p}_{I_s}
  • Isωk{}^{I_s}\boldsymbol{\omega}_k: the auxiliary IMU’s corrected angular velocity at the knot.
  • δtd=td,j−tˉd,j\delta t_d = t_{d,j} - \bar t_{d,j}: only the part of the time offset not yet absorbed. MVIS cuts the auxiliary samples for interval [tk,tk+1][t_k, t_{k+1}] at stamps [tk−tˉd,j, tk+1−tˉd,j][t_k - \bar t_{d,j},\ t_{k+1} - \bar t_{d,j}], with tˉd,j\bar t_{d,j} the starting value. The pose then needs only a small shift, which keeps the first-order propagation accurate even for offsets of tens of ms.

The auxiliary IMU’s own readings are preintegrated over the knot interval, exactly as for the base IMU: corrected by its own biases, intrinsics and t_ga. The residual is the same 15-row residual as the base factor, evaluated on the auxiliary poses above and the auxiliary velocity and biases:

rauxj,k=rimu(IsGTk, IsGTk′, GvIs,k, GvIs,k′, bs,k, bs,k′; xin,s, tga,s)\mathbf{r}_{\text{aux}}^{j,k} = \mathbf{r}_{\text{imu}}\Big( {}^{G}_{I_s}\mathbf{T}_{k},\ {}^{G}_{I_s}\mathbf{T}_{k'},\ {}^{G}\mathbf{v}_{I_s,k},\ {}^{G}\mathbf{v}_{I_s,k'},\ \mathbf{b}_{s,k},\ \mathbf{b}_{s,k'};\ \mathbf{x}_{in,s},\ t_{ga,s}\Big)

with k′k' the next knot. The Jacobians chain the base-factor Jacobians through the pose map above, so they reach the base poses, the base velocities, the extrinsic and td,jt_{d,j}. All are analytic and checked against finite differences.

A point at IpIs{}^{I}\mathbf{p}_{I_s} on a rigid body feels the base acceleration plus two rotation terms:

Isa=IsIR⊤(Ia+[Iω˙]×IpIs+[Iω]×2 IpIs){}^{I_s}\mathbf{a} = {}^{I}_{I_s}\mathbf{R}^\top\Big({}^{I}\mathbf{a} + \big[{}^{I}\dot{\boldsymbol\omega}\big]_\times{}^{I}\mathbf{p}_{I_s} + \big[{}^{I}\boldsymbol\omega\big]_\times^2\,{}^{I}\mathbf{p}_{I_s}\Big)

[⋅]×[\cdot]_\times is the cross-product matrix. With no rotation both terms vanish and the lever arm leaves no trace. The rotation of the extrinsic and td,jt_{d,j} come from the gyro alone: two gyros on one body read the same angular velocity, rotated and shifted in time.

Before the joint solve, MVIS initialises each auxiliary IMU from its gyro and the base gyro alone. Both are direct angular-rate measurements at 100 Hz or more, so this is sharper than going through a camera.

  1. Time offset: scan td,jt_{d,j} and keep the value where the two gyro signals, aligned by the best rotation, agree best. A parabola refines the minimum.
  2. Rotation: the best rotation between the two gyro signals at that offset (the Kabsch method, a closed-form least-squares rotation fit).
  3. Gyro bias from the means of the two signals.
  4. Lever arm: 0. Only the joint solve can see it.

The start is accepted when the minimum is sharp (the error at ±\pm half the search window is more than 1.5 times the minimum) and the two streams overlap by more than 10 s.

  • First knot: velocity and biases, as for the base IMU (0.5 m/s, 0.01 rad/s, 0.1 m/s² by default).
  • Extrinsic and time offset: free. estimate.imu.extrinsics: false or estimate.imu.td: false holds them at their start.
  • Intrinsics: staged as for the base IMU (estimate.imu.intrinsics).
  • t_ga: the same weak 10 ms prior as the base IMU.

Some accelerometers are noisy or interpolated by their driver. Mark such an IMU gyro_only: true and MVIS calibrates it from its gyro alone:

cam_imu.yaml
input:
imus:
- {} # base IMU: always full
- {gyro_only: true} # second IMU: gyro only

The factor keeps only the rotation and gyro-bias rows:

rgyro=[Log ⁣(ΔR⊤  IsGRk⊤ IsGRk′  Exp(δθ)⊤)bg,s,k′−bg,s,k]\mathbf{r}_{\text{gyro}} = \begin{bmatrix} \mathrm{Log}\!\Big(\Delta\mathbf{R}^\top\;{}^{G}_{I_s}\mathbf{R}_k^\top\,{}^{G}_{I_s}\mathbf{R}_{k'}\;\mathrm{Exp}(\delta\boldsymbol\theta)^\top\Big)\\[2pt] \mathbf{b}_{g,s,k'} - \mathbf{b}_{g,s,k} \end{bmatrix}
Gyro-only IMU
Solvedextrinsic rotation, td,jt_{d,j}, gyro bias, and Dw\mathbf{D}_w unless estimate.imu.intrinsics: pinned
Heldlever arm, Da\mathbf{D}_a, Tg\mathbf{T}_g, t_ga (none of them reach the gyro)
Statesthe gyro bias at each knot, no velocity

The base IMU cannot be gyro-only: it needs its accelerometer for gravity and metric scale.

  • Rotate about at least two axes, briskly. The auxiliary rotation and time offset come from angular velocity; the lever arm from angular velocity and angular acceleration.
  • Record long enough for the gyro start: the two IMUs must overlap by more than 10 s.
  • Mount the IMUs rigidly. The model assumes one rigid body.

How we run multi-IMU rigs: step 6: multi-IMU calibration. Why it pays off: Why calibrate IMUs together.