Behaviour cloning is a commonly used strategy for imitation learning and can be extremely effective in constrained domains. However, in cases where the dynamics of an environment may be state dependent and varying, behaviour cloning places a burden on model capacity and the number of demonstrations required. This paper introduces switching density networks, which rely on a categorical reparametrisation for hybrid system identification. This results in a network comprising a classification layer that is followed by a regression layer. We use switching density networks to predict the parameters of hybrid control laws, which are toggled by a switching layer to produce different controller outputs, when conditioned on an input state. This work shows how switching density networks can be used for hybrid system identification in a variety of tasks, successfully identifying the key joint angle goals that make up manipulation tasks, while simultaneously learning image-based goal classifiers and regression networks that predict joint angles from images. We also show that they can cluster the phase space of an inverted pendulum, identifying the balance, spin and pump controllers required to solve this task. Switching density networks can be difficult to train, but we introduce a cross entropy regularisation loss that stabilises training.
Learning from demonstration is an effective method for human users to instruct desired robot behaviour. However, for most non-trivial tasks of practical interest, efficient learning from demonstration depends crucially on inductive bias in the chosen structure for rewards/costs and policies. We address the case where this inductive bias comes from an exchange with a human user. We propose a method in which a learning agent utilizes the information bottleneck layer of a high-parameter variational neural model, with auxiliary loss terms, in order to ground abstract concepts such as spatial relations. The concepts are referred to in natural language instructions and are manifested in the high-dimensional sensory input stream the agent receives from the world. We evaluate the properties of the latent space of the learned model in a photorealistic synthetic environment and particularly focus on examining its usability for downstream tasks. Additionally, through a series of controlled table-top manipulation experiments, we demonstrate that the learned manifold can be used to ground demonstrations as symbolic plans, which can then be executed on a PR2 robot.
Temporally extended and sequenced robot motion tasks are often characterized by discontinuous switches between different types of local dynamics. These change-points can be exploited to build approximate models of the interleaving regions, which in turn allow the design of region-specific controllers. These can then be combined to create the initiation state-space of a final policy. However, such a pipeline can become challenging to implement for combinatorially complex, temporarily extended tasks - especially so when sub-controllers work on different information streams, time scales and action spaces. In this paper, we introduce a method that can compose diverse policies based on scripted motion planning, dynamic motion primitives and neural networks. In order to do this, we extend the options framework to introduce a per-option dynamics module and a global function that evaluates a goal metric. Additionally, we can leverage expert demonstrations to sequence these local policies, converting the learning problem in hierarchical reinforcement learning to a planning problem at inference time. We first illustrate the core concepts with an MDP benchmark, and then with a physical gear assembly task solved on a PR2 robot. We show that the proposed approach successfully discovers the optimal sequence of policies and solves both tasks efficiently.
Robots performing tasks in dynamic environments would benefit greatly from understanding the underlying environment motion, in order to make future predictions and to synthesize effective control policies that use this inductive bias. Online system identification is therefore a fundamental requirement for robust autonomous agents. When the dynamics involves multiple modes (due to contacts or interactions between objects), and when system identification must proceed directly from a rich sensory stream such as video, then traditional methods for system identification may not be well suited. We propose an approach wherein fast parameter estimation with a model can be seamlessly combined with a recurrent variational autoencoder. Our Physics-based recurrent variational autoencoder model includes an additional loss that enforces conformity with the structure of a physically based dynamics model. This enables the resulting model to encode parameters such as position, velocity, restitution, air drag and other physical properties of the system. The model can be trained entirely in simulation, in an end-to-end manner with domain randomization, to perform online system identification, and probabilistic forward predictions of parameters of interest. We benchmark against existing system identification methods and demonstrate that Vid2Param outperforms the baselines in terms of speed and accuracy of identification, and also provides uncertainty quantification in the form of a distribution over future trajectories. Furthermore, we illustrate the utility of this in physical experiments wherein a PR2 robot with velocity constrained arm must intercept a bouncing ball, by estimating the physical parameters of this ball directly from the video trace after the ball is released.
We aim to perform unsupervised discovery of objects and their states such as location and velocity, as well as physical system parameters such as mass and gravity from video -- given only the differential equations governing the scene dynamics. Existing physical scene understanding methods require either object state supervision, or do not integrate with differentiable physics to learn interpretable system parameters and states. We address this problem through a $\textit{physics-as-inverse-graphics}$ approach that brings together vision-as-inverse-graphics and differentiable physics engines. This framework allows us to perform long term extrapolative video prediction, as well as vision-based model-predictive control. Our approach significantly outperforms related unsupervised methods in long-term future frame prediction of systems with interacting objects (such as ball-spring or 3-body gravitational systems). We further show the value of this tight vision-physics integration by demonstrating data-efficient learning of vision-actuated model-based control for a pendulum system. The controller's interpretability also provides unique capabilities in goal-driven control and physical reasoning for zero-data adaptation.
Convolutional Neural Networks (CNNs) have been used successfully across a broad range of areas including data mining, object detection, and in business. The dominance of CNNs follows a breakthrough by Alex Krizhevsky which showed improvements by dramatically reducing the error rate obtained in a general image classification task from 26.2% to 15.4%. In road safety, CNNs have been applied widely to the detection of traffic signs, obstacle detection, and lane departure checking. In addition, CNNs have been used in data mining systems that monitor driving patterns and recommend rest breaks when appropriate. This paper presents a driver drowsiness detection system and shows that there are potential social challenges regarding the application of these techniques, by highlighting problems in detecting dark-skinned driver's faces. This is a particularly important challenge in African contexts, where there are more dark-skinned drivers. Unfortunately, publicly available datasets are often captured in different cultural contexts, and therefore do not cover all ethnicities, which can lead to false detections or racially biased models. This work evaluates the performance obtained when training convolutional neural network models on commonly used driver drowsiness detection datasets and testing on datasets specifically chosen for broader representation. Results show that models trained using publicly available datasets suffer extensively from over-fitting, and can exhibit racial bias, as shown by testing on a more representative dataset. We propose a novel visualisation technique that can assist in identifying groups of people where there might be the potential of discrimination, using Principal Component Analysis (PCA) to produce a grid of faces sorted by similarity, and combining these with a model accuracy overlay.
A great deal of work aims to discover general purpose models of image interest or memorability for visual search and information retrieval. This paper argues that image interest is often domain and user specific, and that mechanisms for learning about this domain-specific image interest as quickly as possible, while limiting the amount of data-labelling required, are often more useful to end-users. Specifically, this paper is concerned with the small to medium-sized data regime regularly faced by practising data scientists, who are often required to build turnkey models for end-users with domain-specific challenges. This work uses pairwise image comparisons to reduce the labelling burden on these users, and shows that Gaussian process smoothing in image feature space can be used to build probabilistic models of image interest extremely quickly for a wide range of problems, and performs similarly to recent deep learning approaches trained using pairwise ranking losses. The Gaussian process model used in this work interpolates image interest inferred using a Bayesian ranking approach over image features extracted using a pre-trained convolutional neural network. This probabilistic approach produces image interests paired with uncertainties that can be used to identify images for which additional labelling is required and measure inference convergence. Results obtained on five distinct datasets reinforce recent findings that pre-trained convolutional neural networks can be used to extract useful representations applicable across multiple domains, and highlight the fact that domain-specific image interest does not always correlate with concepts like image memorability.
Recent work has shown significant progress in the direction of synthetic data generation using Generative Adversarial Networks (GANs). GANs have been applied in many fields of computer vision including text-to-image conversion, domain transfer, super-resolution, and image-to-video applications. In computer vision, traditional GANs are based on deep convolutional neural networks. However, deep convolutional neural networks can require extensive computational resources because they are based on multiple operations performed by convolutional layers, which can consist of millions of trainable parameters. Training a GAN model can be difficult and it takes a significant amount of time to reach an equilibrium point. In this paper, we investigate the use of depthwise separable convolutions to reduce training time while maintaining data generation performance. Our results show that a DepthwiseGAN architecture can generate realistic images in shorter training periods when compared to a StarGan architecture, but that model capacity still plays a significant role in generative modelling. In addition, we show that depthwise separable convolutions perform best when only applied to the generator. For quality evaluation of generated images, we use the Fr\'echet Inception Distance (FID), which compares the similarity between the generated image distribution and that of the training dataset.