Actuation Subspace Prediction with Neural Householder Transforms

Abstract

Choosing an appropriate action representation is an integral part of solving robotic manipulation problems. Published approaches include latent action models, which train context-conditioned neural networks to map lowdimensional latent actions to high-dimensional actuation commands. Such models can have a large number of parameters and can be difficult to interpret from a user’s perspective. In this thesis, we propose that similar performance gains in robotics tasks can be achieved by restructuring the neural network to map observations to a basis for a context-dependent linear actuation subspace. This results in an action interface wherein a user’s actions determine a linear combination of state-conditioned actuation basis vectors. We introduce the Neural Householder Transform (NHT) as a method for computing this basis. This thesis describes the development of NHT: from computing an unconstrained basis for a state-conditioned linear map (SCL), to computing an orthonormal basis that changes smoothly with respect to the input context. Two teleoperation user studies indicated that participants preferred using SCL, and tend to achieve higher completion rates with SCL compared to latent action models. In addition, simulation results showed that reinforcement learning agents trained with NHT in kinematic manipulation and locomotion environments tend to be more robust to hyperparameter choice and achieve higher final success rates compared to agents trained with alternative action representations.