PVEs: Position-Velocity Encoders for Unsupervised Learning of Structured State Representations
This addresses the challenge of unsupervised representation learning for robotics and control, though it appears incremental as it builds on prior work with physical priors.
The authors tackled the problem of learning structured state representations from images without supervision by proposing position-velocity encoders (PVEs), which encode images into positions and velocities of objects using physical priors instead of reconstruction, achieving promising preliminary results on simulated control tasks.
We propose position-velocity encoders (PVEs) which learn---without supervision---to encode images to positions and velocities of task-relevant objects. PVEs encode a single image into a low-dimensional position state and compute the velocity state from finite differences in position. In contrast to autoencoders, position-velocity encoders are not trained by image reconstruction, but by making the position-velocity representation consistent with priors about interacting with the physical world. We applied PVEs to several simulated control tasks from pixels and achieved promising preliminary results.