Masked Proprioception Autoencoders for Humanoid Locomotion: Self-Supervised Learning to Reconstruct IMU Signals

This post explores a Masked Autoencoder (MAE) approach applied to proprioceptive signals — specifically IMU data from humanoid locomotion rollouts.

Inspired by masked image modeling, the idea is to mask a subset of IMU measurements across time and train a transformer encoder-decoder to reconstruct the missing signals. This forces the model to learn rich temporal and kinematic structure from locomotion data in a fully self-supervised manner, without any labels.

Read the full write-up on Medium: Masked Proprioception Autoencoders for Humanoid Locomotion