We present DemoStart, a novel auto-curriculum reinforcement learning method capable of learning complex manipulation behaviors on an arm equipped with a three- fingered robotic hand, from only a sparse reward and a handful of demonstrations in simulation. Learning from simulation drastically reduces the development cycle of behavior generation, and domain randomization techniques are leveraged to achieve successful zero-shot sim-to- real transfer. Transferred policies are learned directly from raw pixels from multiple cameras and robot proprioception. Our approach outperforms policies learned from demonstrations on the real robot and requires 100 times fewer demonstrations, collected in simulation. More details and videos in sites.google.com/view/demostart.

Learning based methods for dexterous manipulation have made notable progress in recent years, and they can now produce solutions to complex tasks. However, learned policies often still lack reliability and exhibit limited robustness to important factors of variation. One failure pattern that can be observed across many settings is that policies idle, i.e. they cease to move beyond a small region of states, often indefinitely, when they reach certain states. This policy idling is often a reflection of the training data. For instance, it can occur when the data contains small actions in areas where the robot needs to perform high-precision motions, e.g., when preparing to grasp an object or object insertion. Prior works have tried to mitigate this phenomenon e.g. by filtering the training data or modifying the control frequency. However, these approaches can negatively impact policy performance in other ways. As an alternative, we investigate how to leverage the detectability of idling behavior to inform exploration and policy improvement. Our approach, Pause-Induced Perturbations (PIP), applies perturbations at detected idling states, thus helping it to escape problematic basins of attraction. On a range of challenging simulated dual-arm tasks, we find that this simple approach can already noticeably improve test-time performance, with no additional supervision or training. Furthermore, since the robot tends to idle at critical points in a movement, we also find that learning from the resulting episodes leads to better iterative policy improvement compared to prior approaches. Our perturbation strategy also leads to a 15-35% improvement in absolute success rate on a real-world insertion task that requires complex multi-finger manipulation.

We introduce Hindsight Off-policy Options (HO2), a data-efficient option learning algorithm. Given any trajectory, HO2 infers likely option choices and backpropagates through the dynamic programming inference procedure to robustly train all policy components off-policy and end-to-end. The approach outperforms existing option learning methods on common benchmarks. To better understand the option framework and disentangle benefits from both temporal and action abstraction, we evaluate ablations with flat policies and mixture policies with comparable optimization. The results highlight the importance of both types of abstraction as well as off-policy training and trust-region constraints, particularly in challenging, simulated 3D robot manipulation tasks from raw pixel inputs. Finally, we intuitively adapt the inference step to investigate the effect of increased temporal abstraction on training with pre-trained options and from scratch.

Continual learning aims to improve the ability of modern learning systems to deal with non-stationary distributions, typically by attempting to learn a series of tasks sequentially. Prior art in the field has largely considered supervised or reinforcement learning tasks, and often assumes full knowledge of task labels and boundaries. In this work, we propose an approach (CURL) to tackle a more general problem that we will refer to as unsupervised continual learning. The focus is on learning representations without any knowledge about task identity, and we explore scenarios when there are abrupt changes between tasks, smooth transitions from one task to another, or even when the data is shuffled. The proposed approach performs task inference directly within the model, is able to dynamically expand to capture new concepts over its lifetime, and incorporates additional rehearsal-based techniques to deal with catastrophic forgetting. We demonstrate the efficacy of CURL in an unsupervised learning setting with MNIST and Omniglot, where the lack of labels ensures no information is leaked about the task. Further, we demonstrate strong performance compared to prior art in an i.i.d setting, or when adapting the technique to supervised tasks such as incremental class learning.

This paper proposes a computationally efficient approach to detecting objects natively in 3D point clouds using convolutional neural networks (CNNs). In particular, this is achieved by leveraging a feature-centric voting scheme to implement novel convolutional layers which explicitly exploit the sparsity encountered in the input. To this end, we examine the trade-off between accuracy and speed for different architectures and additionally propose to use an L1penalty on the filter activations to further encourage sparsity in the intermediate representations. To the best of our knowledge, this is the first work to propose sparse convolutional layers and L1regularisation for efficient large-scale processing of 3D data. We demonstrate the efficacy of our approach on the KITTI object detection benchmark and show that Vote3Deep models with as few as three layers outperform the previous state of the art in both laser and laser-vision based approaches by margins of up to 40% while remaining highly competitive in terms of processing time.

Autonomous underwater vehicles (AUVs) are widely used to perform information gathering missions in unseen environments. Given the sheer size of the ocean environment, and the time and energy constraints of an AUV, it is important to consider the potential utility of candidate missions when performing survey planning. In this paper, we utilise a multimodal learning approach to capture the relationship between in-situ visual observations, and shipborne bathymetry (ocean depth) data that are freely available a priori. We then derive information-theoretic measures under this model that predict the amount of visual information gain at an unobserved location based on the bathymetric features. Unlike previous approaches, these measures consider the value of additional visual features, rather than just the habitat labels obtained. Experimental results with a toy dataset and real marine data demonstrate that the approach can be used to predict the true utility of unexplored areas.

Autonomous Underwater Vehicles (AUVs) gather large volumes of visual imagery, which can help monitor marine ecosystems and plan future surveys. One key task in marine ecology is benthic habitat mapping, the classification of large regions of the ocean floor into broad habitat categories. Since visual data only covers a small fraction of the ocean floor, traditional habitat mapping is performed using shipborne acoustic multi-beam data, with visual data as ground truth. However, given the high resolution and rich textural cues in visual data, an ideal approach should explicitly utilise visual features in the classification process. To this end, we propose a multimodal model which utilises visual data and shipborne multi-beam bathymetry to perform both classification and sampling tasks. Our algorithm learns the relationship between both modalities, but is also effective when visual data is missing. Our results suggest that by performing multimodal learning, classification performance is improved in scenarios where visual data is unavailable, such as the habitat mapping scenario. We also demonstrate empirically that the model is able to perform generative tasks, producing plausible samples from the underlying data-generating distribution.

Existing approaches to visual Simultaneous Localization and Mapping (SLAM) typically utilize points as visual feature primitives to represent landmarks in the environment. Since these techniques mostly use image points from a standard feature point detector, they do not explicitly map objects or regions of interest. Our work is motivated by the need for different SLAM techniques in path and riverine settings, where feature points can be scarce or may not adequately represent the environment. Accordingly, the proposed approach uses cubic Bézier curves as stereo vision primitives and offers a novel SLAM formulation to update the curve parameters and vehicle pose. This method eliminates the need for point-based stereo matching, with an optimization procedure to directly extract the curve information in the world frame from noisy edge measurements. Further, the proposed algorithm enables navigation with fewer feature states than most point-based techniques, and is able to produce a map which only provides detail in key areas. Results in simulation and with vision data validate that the proposed method can be effective in estimating the 6DOF pose of the stereo camera, and can produce structured, uncluttered maps.