I had a look on the XSens website and following a few references lists to find:

Schepers, Guiberti & Bellusci (2018) Xsens MVN: Consistent Tracking of Human Motion Using Inertial Sensing, version 2018.0. XSENS Technologies
Picerno, Cereatti, Sappozzo (2008) Joint Kinematics estimates using wearable inertial and magnetic sensing modules. Gait & Posture, 28, 588-595
van den Noort, Ferrari, Cutti, Becher, Harlaar (2013) Gait Analysis in Children with Cerebral Palsy via inertial and magnetic sensors. Med Biol Eng Comput, 51, 377-386
Al-Amri, Nicholas, Button, Sparkes, Sheeran, Davies (2018) Inertia Measurement Units for Clinical Movement Analysis: Reliablity and Concurrant Validity. Sensors, 18, 719.
Cloete & Sheffer (2010) Repeatability of an off the shelf, full body inertial motion capture system during clinical gait analysis. 32nd Annual International Congress of the IEEE EMBS, Argentina.

The first four assess concurrent validity (agreement) between IMU (XSens) and Optical (OpenSim, CAST, PiG). All find no agreement in all gait joint angles produced by IMU and Optical measurement systems with huge variability across all lower limb joint angles (4.2 < StDev < 17.5 degs). The results of Al-Amri were so poor for non-sagital rotations that aCMC was negative (CMC <0.5) and not reported. These studies artificially increase agreement by removing all offsets from gait curves but still produce large differences between all joint angle curves (2.5 < StDev < 6.1). These are still larger than inter-subject variability (1.2 < StDev < 4.1) and considerably larger than intra-session variability (0.55 < StDev < 0.95) expected of optical cluster based approach with no offsets removed. Of concern is that Picerno etal. and van den Noort et al. placed markers directly on each IMU housing to obtain a direct comparison of the agreement between the two measurement system in reconstructing the same axes systems. Although it should be noted that OpenSimm, PiG and CAST with poor validly and reliably and inability to describe gait non-sagittal rotations of the knee and ankle as well as transverse rotations of the hip, are not a good choices for concurrent validity.

Cloete etal examined inter-session reliability of IMU (Xsens), optical and electromagnetic systems. All systems produced similar inter-session reliability for all gait joint angles, with no one system standing out. This may be partly due to offsets being removed from gait joint curves. It also means that the actual inter-session reliability is unknown. With offsets removed the three systems produced fair inter-session reliability (1.1 < StDev < 2.2), but still larger than expected for optical cluster based methods (0.55 < StDev < 0.95). No details are given of the optical system or approach used. The inter-session reliability of Al-Amri was poor for sagittal including hip Abd/Add (1.1 < StDev < 2.3) and particular poor for non-sagital (2.0 < StDev < 4.0) joint angles.

Despite the poor results and validity of joint angle data not been addresses, they all conclude the IMU methods are valid, reliable and able to produce 3D gait joint angles in research or clinical setting.