We propose Scheduled Auxiliary Control (SAC-X), a new learning paradigm in the context of Reinforcement Learning (RL). SAC-X enables learning of complex behaviors - from scratch - in the presence of multiple sparse reward signals. To this end, the agent is equipped with a set of general auxiliary tasks, that it attempts to learn simultaneously via off-policy RL. The key idea behind our method is that active (learned) scheduling and execution of auxiliary policies allows the agent to efficiently explore its environment - enabling it to excel at sparse reward RL. Our experiments in several challenging robotic manipulation settings demonstrate the power of our approach.

We present a novel model architecture which leverages deep learning tools to perform exact Bayesian inference on sets of high dimensional, complex observations. Our model is provably exchangeable, meaning that the joint distribution over observations is invariant under permutation: this property lies at the heart of Bayesian inference. The model does not require variational approximations to train, and new samples can be generated conditional on previous samples, with cost linear in the size of the conditioning set. The advantages of our architecture are demonstrated on learning tasks that require generalisation from short observed sequences while modelling sequence variability, such as conditional image generation, few-shot learning, and anomaly detection.

Incorporating the body dynamics of compliant robots into their controller architectures can drastically reduce the complexity of locomotion control. An extreme version of this embodied control principle was demonstrated in highly compliant tensegrity robots, for which stable gait generation was achieved by using only optimized linear feedback from the robot's sensors to its actuators. The morphology of quadrupedal robots has previously been used for sensing and for control of a compliant spine, but never for gait generation. In this paper, we successfully apply embodied control to the compliant, quadrupedal Oncilla robot. As initial experiments indicated that mere linear feedback does not suffice, we explore the minimal requirements for robust gait generation in terms of memory and nonlinear complexity. Our results show that a memoryless feedback controller can generate a stable trot by learning the desired nonlinear relation between the input and the output signals. We believe this method can provide a very useful tool for transferring knowledge from open loop to closed loop control on compliant robots.

Using data retrieved from the Puppy II robot at the University of Zurich (UZH), we show that machine learning techniques with non-linearities and fading memory are effective for terrain classification, both supervised and unsupervised, even with a limited selection of input sensors. The results indicate that most information for terrain classification is found in the combination of tactile sensors and proprioceptive joint angle sensors. The classification error is small enough to have a robot adapt the gait to the terrain and hence move more robustly.