Niko Sünderhauf

Matt Deitke, Dhruv Batra, Yonatan Bisk et al. · allen-ai, cmu

We present a retrospective on the state of Embodied AI research. Our analysis focuses on 13 challenges presented at the Embodied AI Workshop at CVPR. These challenges are grouped into three themes: (1) visual navigation, (2) rearrangement, and (3) embodied vision-and-language. We discuss the dominant datasets within each theme, evaluation metrics for the challenges, and the performance of state-of-the-art models. We highlight commonalities between top approaches to the challenges and identify potential future directions for Embodied AI research.

Niko Sünderhauf, Jad Abou-Chakra, Dimity Miller

We show that ensembling effectively quantifies model uncertainty in Neural Radiance Fields (NeRFs) if a density-aware epistemic uncertainty term is considered. The naive ensembles investigated in prior work simply average rendered RGB images to quantify the model uncertainty caused by conflicting explanations of the observed scene. In contrast, we additionally consider the termination probabilities along individual rays to identify epistemic model uncertainty due to a lack of knowledge about the parts of a scene unobserved during training. We achieve new state-of-the-art performance across established uncertainty quantification benchmarks for NeRFs, outperforming methods that require complex changes to the NeRF architecture and training regime. We furthermore demonstrate that NeRF uncertainty can be utilised for next-best view selection and model refinement.

Krishan Rana, Ming Xu, Brendan Tidd et al.

Skill-based reinforcement learning (RL) has emerged as a promising strategy to leverage prior knowledge for accelerated robot learning. Skills are typically extracted from expert demonstrations and are embedded into a latent space from which they can be sampled as actions by a high-level RL agent. However, this skill space is expansive, and not all skills are relevant for a given robot state, making exploration difficult. Furthermore, the downstream RL agent is limited to learning structurally similar tasks to those used to construct the skill space. We firstly propose accelerating exploration in the skill space using state-conditioned generative models to directly bias the high-level agent towards only sampling skills relevant to a given state based on prior experience. Next, we propose a low-level residual policy for fine-grained skill adaptation enabling downstream RL agents to adapt to unseen task variations. Finally, we validate our approach across four challenging manipulation tasks that differ from those used to build the skill space, demonstrating our ability to learn across task variations while significantly accelerating exploration, outperforming prior works. Code and videos are available on our project website: https://krishanrana.github.io/reskill.

Samuel Wilson, Tobias Fischer, Feras Dayoub et al.

We address the problem of out-of-distribution (OOD) detection for the task of object detection. We show that residual convolutional layers with batch normalisation produce Sensitivity-Aware FEatures (SAFE) that are consistently powerful for distinguishing in-distribution from out-of-distribution detections. We extract SAFE vectors for every detected object, and train a multilayer perceptron on the surrogate task of distinguishing adversarially perturbed from clean in-distribution examples. This circumvents the need for realistic OOD training data, computationally expensive generative models, or retraining of the base object detector. SAFE outperforms the state-of-the-art OOD object detectors on multiple benchmarks by large margins, e.g. reducing the FPR95 by an absolute 30.6% from 48.3% to 17.7% on the OpenImages dataset.

Jad Abou-Chakra, Feras Dayoub, Niko Sünderhauf

While existing Neural Radiance Fields (NeRFs) for dynamic scenes are offline methods with an emphasis on visual fidelity, our paper addresses the online use case that prioritises real-time adaptability. We present ParticleNeRF, a new approach that dynamically adapts to changes in the scene geometry by learning an up-to-date representation online, every 200ms. ParticleNeRF achieves this using a novel particle-based parametric encoding. We couple features to particles in space and backpropagate the photometric reconstruction loss into the particles' position gradients, which are then interpreted as velocity vectors. Governed by a lightweight physics system to handle collisions, this lets the features move freely with the changing scene geometry. We demonstrate ParticleNeRF on various dynamic scenes containing translating, rotating, articulated, and deformable objects. ParticleNeRF is the first online dynamic NeRF and achieves fast adaptability with better visual fidelity than brute-force online InstantNGP and other baseline approaches on dynamic scenes with online constraints. Videos of our system can be found at our project website https://sites.google.com/view/particlenerf.

David Pershouse, Feras Dayoub, Dimity Miller et al.

We address the challenging problem of open world object detection (OWOD), where object detectors must identify objects from known classes while also identifying and continually learning to detect novel objects. Prior work has resulted in detectors that have a relatively low ability to detect novel objects, and a high likelihood of classifying a novel object as one of the known classes. We approach the problem by identifying the three main challenges that OWOD presents and introduce OW-RCNN, an open world object detector that addresses each of these three challenges. OW-RCNN establishes a new state of the art using the open-world evaluation protocol on MS-COCO, showing a drastically increased ability to detect novel objects (16-21% absolute increase in U-Recall), to avoid their misclassification as one of the known classes (up to 52% reduction in A-OSE), and to incrementally learn to detect them while maintaining performance on previously known classes (1-6% absolute increase in mAP).

Krishan Rana, Andrew Melnik, Niko Sünderhauf

In this paper, we introduce a method for unifying language, action, and state information in a shared embedding space to facilitate a range of downstream tasks in robot learning. Our method, Contrastive Language, Action, and State Pre-training (CLASP), extends the CLIP formulation by incorporating distributional learning, capturing the inherent complexities and one-to-many relationships in behaviour-text alignment. By employing distributional outputs for both text and behaviour encoders, our model effectively associates diverse textual commands with a single behaviour and vice-versa. We demonstrate the utility of our method for the following downstream tasks: zero-shot text-behaviour retrieval, captioning unseen robot behaviours, and learning a behaviour prior for language-conditioned reinforcement learning. Our distributional encoders exhibit superior retrieval and captioning performance on unseen datasets, and the ability to generate meaningful exploratory behaviours from textual commands, capturing the intricate relationships between language, action, and state. This work represents an initial step towards developing a unified pre-trained model for robotics, with the potential to generalise to a broad range of downstream tasks.

Federico Ceola, Lorenzo Natale, Niko Sünderhauf et al.

Instructing a robot to complete an everyday task within our homes has been a long-standing challenge for robotics. While recent progress in language-conditioned imitation learning and offline reinforcement learning has demonstrated impressive performance across a wide range of tasks, they are typically limited to short-horizon tasks -- not reflective of those a home robot would be expected to complete. While existing architectures have the potential to learn these desired behaviours, the lack of the necessary long-horizon, multi-step datasets for real robotic systems poses a significant challenge. To this end, we present the Long-Horizon Manipulation (LHManip) dataset comprising 200 episodes, demonstrating 20 different manipulation tasks via real robot teleoperation. The tasks entail multiple sub-tasks, including grasping, pushing, stacking and throwing objects in highly cluttered environments. Each task is paired with a natural language instruction and multi-camera viewpoints for point-cloud or NeRF reconstruction. In total, the dataset comprises 176,278 observation-action pairs which form part of the Open X-Embodiment dataset. The full LHManip dataset is made publicly available at https://github.com/fedeceola/LHManip.

Chamuditha Jayanga Galappaththige, Jason Lai, Lloyd Windrim et al.

Online Scene Change Detection (SCD) is an extremely challenging problem that requires an agent to detect relevant changes on the fly while observing the scene from unconstrained viewpoints. Existing online SCD methods are significantly less accurate than offline approaches. We present the first online SCD approach that is pose-agnostic, label-free, and ensures multi-view consistency, while operating at over 10 FPS and achieving new state-of-the-art performance, surpassing even the best offline approaches. Our method introduces a new self-supervised fusion loss to infer scene changes from multiple cues and observations, PnP-based fast pose estimation against the reference scene, and a fast change-guided update strategy for the 3D Gaussian Splatting scene representation. Extensive experiments on complex real-world datasets demonstrate that our approach outperforms both online and offline baselines.

Jesse Haviland, Niko Sünderhauf, Peter Corke

We present the design and implementation of a taskable reactive mobile manipulation system. In contrary to related work, we treat the arm and base degrees of freedom as a holistic structure which greatly improves the speed and fluidity of the resulting motion. At the core of this approach is a robust and reactive motion controller which can achieve a desired end-effector pose, while avoiding joint position and velocity limits, and ensuring the mobile manipulator is manoeuvrable throughout the trajectory. This can support sensor-based behaviours such as closed-loop visual grasping. As no planning is involved in our approach, the robot is never stationary thinking about what to do next. We show the versatility of our holistic motion controller by implementing a pick and place system using behaviour trees and demonstrate this task on a 9-degree-of-freedom mobile manipulator. Additionally, we provide an open-source implementation of our motion controller for both non-holonomic and omnidirectional mobile manipulators available at jhavl.github.io/holistic.

Haoyang Zhang, Ying Wang, Feras Dayoub et al.

Do you want to improve 1.0 AP for your object detector without any inference cost and any change to your detector? Let us tell you such a recipe. It is surprisingly simple: train your detector for an extra 12 epochs using cyclical learning rates and then average these 12 checkpoints as your final detection model}. This potent recipe is inspired by Stochastic Weights Averaging (SWA), which is proposed in arXiv:1803.05407 for improving generalization in deep neural networks. We found it also very effective in object detection. In this technique report, we systematically investigate the effects of applying SWA to object detection as well as instance segmentation. Through extensive experiments, we discover the aforementioned workable policy of performing SWA in object detection, and we consistently achieve $\sim$1.0 AP improvement over various popular detectors on the challenging COCO benchmark, including Mask RCNN, Faster RCNN, RetinaNet, FCOS, YOLOv3 and VFNet. We hope this work will make more researchers in object detection know this technique and help them train better object detectors. Code is available at: https://github.com/hyz-xmaster/swa_object_detection .

Haoyang Zhang, Ying Wang, Feras Dayoub et al.

Accurately ranking the vast number of candidate detections is crucial for dense object detectors to achieve high performance. Prior work uses the classification score or a combination of classification and predicted localization scores to rank candidates. However, neither option results in a reliable ranking, thus degrading detection performance. In this paper, we propose to learn an Iou-aware Classification Score (IACS) as a joint representation of object presence confidence and localization accuracy. We show that dense object detectors can achieve a more accurate ranking of candidate detections based on the IACS. We design a new loss function, named Varifocal Loss, to train a dense object detector to predict the IACS, and propose a new star-shaped bounding box feature representation for IACS prediction and bounding box refinement. Combining these two new components and a bounding box refinement branch, we build an IoU-aware dense object detector based on the FCOS+ATSS architecture, that we call VarifocalNet or VFNet for short. Extensive experiments on MS COCO show that our VFNet consistently surpasses the strong baseline by $\sim$2.0 AP with different backbones. Our best model VFNet-X-1200 with Res2Net-101-DCN achieves a single-model single-scale AP of 55.1 on COCO test-dev, which is state-of-the-art among various object detectors.Code is available at https://github.com/hyz-xmaster/VarifocalNet .

Dimity Miller, Niko Sünderhauf, Alex Kenna et al.

Are vision-language models (VLMs) for open-vocabulary perception inherently open-set models because they are trained on internet-scale datasets? We answer this question with a clear no - VLMs introduce closed-set assumptions via their finite query set, making them vulnerable to open-set conditions. We systematically evaluate VLMs for open-set recognition and find they frequently misclassify objects not contained in their query set, leading to alarmingly low precision when tuned for high recall and vice versa. We show that naively increasing the size of the query set to contain more and more classes does not mitigate this problem, but instead causes diminishing task performance and open-set performance. We establish a revised definition of the open-set problem for the age of VLMs, define a new benchmark and evaluation protocol to facilitate standardised evaluation and research in this important area, and evaluate promising baseline approaches based on predictive uncertainty and dedicated negative embeddings on a range of open-vocabulary VLM classifiers and object detectors.

Sourav Garg, Krishan Rana, Mehdi Hosseinzadeh et al.

Mapping is crucial for spatial reasoning, planning and robot navigation. Existing approaches range from metric, which require precise geometry-based optimization, to purely topological, where image-as-node based graphs lack explicit object-level reasoning and interconnectivity. In this paper, we propose a novel topological representation of an environment based on "image segments", which are semantically meaningful and open-vocabulary queryable, conferring several advantages over previous works based on pixel-level features. Unlike 3D scene graphs, we create a purely topological graph with segments as nodes, where edges are formed by a) associating segment-level descriptors between pairs of consecutive images and b) connecting neighboring segments within an image using their pixel centroids. This unveils a "continuous sense of a place", defined by inter-image persistence of segments along with their intra-image neighbours. It further enables us to represent and update segment-level descriptors through neighborhood aggregation using graph convolution layers, which improves robot localization based on segment-level retrieval. Using real-world data, we show how our proposed map representation can be used to i) generate navigation plans in the form of "hops over segments" and ii) search for target objects using natural language queries describing spatial relations of objects. Furthermore, we quantitatively analyze data association at the segment level, which underpins inter-image connectivity during mapping and segment-level localization when revisiting the same place. Finally, we show preliminary trials on segment-level `hopping' based zero-shot real-world navigation. Project page with supplementary details: oravus.github.io/RoboHop/

Samuel Wilson, Tobias Fischer, Niko Sünderhauf et al.

We introduce powerful ideas from Hyperdimensional Computing into the challenging field of Out-of-Distribution (OOD) detection. In contrast to most existing work that performs OOD detection based on only a single layer of a neural network, we use similarity-preserving semi-orthogonal projection matrices to project the feature maps from multiple layers into a common vector space. By repeatedly applying the bundling operation $\oplus$, we create expressive class-specific descriptor vectors for all in-distribution classes. At test time, a simple and efficient cosine similarity calculation between descriptor vectors consistently identifies OOD samples with better performance than the current state-of-the-art. We show that the hyperdimensional fusion of multiple network layers is critical to achieve best general performance.

Krishan Rana, Vibhavari Dasagi, Jesse Haviland et al.

While deep reinforcement learning (RL) agents have demonstrated incredible potential in attaining dexterous behaviours for robotics, they tend to make errors when deployed in the real world due to mismatches between the training and execution environments. In contrast, the classical robotics community have developed a range of controllers that can safely operate across most states in the real world given their explicit derivation. These controllers however lack the dexterity required for complex tasks given limitations in analytical modelling and approximations. In this paper, we propose Bayesian Controller Fusion (BCF), a novel uncertainty-aware deployment strategy that combines the strengths of deep RL policies and traditional handcrafted controllers. In this framework, we can perform zero-shot sim-to-real transfer, where our uncertainty based formulation allows the robot to reliably act within out-of-distribution states by leveraging the handcrafted controller while gaining the dexterity of the learned system otherwise. We show promising results on two real-world continuous control tasks, where BCF outperforms both the standalone policy and controller, surpassing what either can achieve independently. A supplementary video demonstrating our system is provided at https://bit.ly/bcf_deploy.

Suman Raj Bista, David Hall, Ben Talbot et al.

Recent Semantic SLAM methods combine classical geometry-based estimation with deep learning-based object detection or semantic segmentation. In this paper we evaluate the quality of semantic maps generated by state-of-the-art class- and instance-aware dense semantic SLAM algorithms whose codes are publicly available and explore the impacts both semantic segmentation and pose estimation have on the quality of semantic maps. We obtain these results by providing algorithms with ground-truth pose and/or semantic segmentation data available from simulated environments. We establish that semantic segmentation is the largest source of error through our experiments, dropping mAP and OMQ performance by up to 74.3% and 71.3% respectively.

Quazi Marufur Rahman, Niko Sünderhauf, Peter Corke et al.

Semantic segmentation is an important task that helps autonomous vehicles understand their surroundings and navigate safely. During deployment, even the most mature segmentation models are vulnerable to various external factors that can degrade the segmentation performance with potentially catastrophic consequences for the vehicle and its surroundings. To address this issue, we propose a failure detection framework to identify pixel-level misclassification. We do so by exploiting internal features of the segmentation model and training it simultaneously with a failure detection network. During deployment, the failure detector can flag areas in the image where the segmentation model have failed to segment correctly. We evaluate the proposed approach against state-of-the-art methods and achieve 12.30%, 9.46%, and 9.65% performance improvement in the AUPR-Error metric for Cityscapes, BDD100K, and Mapillary semantic segmentation datasets.

Krishan Rana, Vibhavari Dasagi, Jesse Haviland et al.

We present Bayesian Controller Fusion (BCF): a hybrid control strategy that combines the strengths of traditional hand-crafted controllers and model-free deep reinforcement learning (RL). BCF thrives in the robotics domain, where reliable but suboptimal control priors exist for many tasks, but RL from scratch remains unsafe and data-inefficient. By fusing uncertainty-aware distributional outputs from each system, BCF arbitrates control between them, exploiting their respective strengths. We study BCF on two real-world robotics tasks involving navigation in a vast and long-horizon environment, and a complex reaching task that involves manipulability maximisation. For both these domains, simple handcrafted controllers exist that can solve the task at hand in a risk-averse manner but do not necessarily exhibit the optimal solution given limitations in analytical modelling, controller miscalibration and task variation. As exploration is naturally guided by the prior in the early stages of training, BCF accelerates learning, while substantially improving beyond the performance of the control prior, as the policy gains more experience. More importantly, given the risk-aversity of the control prior, BCF ensures safe exploration and deployment, where the control prior naturally dominates the action distribution in states unknown to the policy. We additionally show BCF's applicability to the zero-shot sim-to-real setting and its ability to deal with out-of-distribution states in the real world. BCF is a promising approach towards combining the complementary strengths of deep RL and traditional robotic control, surpassing what either can achieve independently. The code and supplementary video material are made publicly available at https://krishanrana.github.io/bcf.

Ming Xu, Tobias Fischer, Niko Sünderhauf et al.

Probabilistic state-estimation approaches offer a principled foundation for designing localization systems, because they naturally integrate sequences of imperfect motion and exteroceptive sensor data. Recently, probabilistic localization systems utilizing appearance-invariant visual place recognition (VPR) methods as the primary exteroceptive sensor have demonstrated state-of-the-art performance in the presence of substantial appearance change. However, existing systems 1) do not fully utilize odometry data within the motion models, and 2) are unable to handle route deviations, due to the assumption that query traverses exactly repeat the mapping traverse. To address these shortcomings, we present a new probabilistic topometric localization system which incorporates full 3-dof odometry into the motion model and furthermore, adds an "off-map" state within the state-estimation framework, allowing query traverses which feature significant route detours from the reference map to be successfully localized. We perform extensive evaluation on multiple query traverses from the Oxford RobotCar dataset exhibiting both significant appearance change and deviations from routes previously traversed. In particular, we evaluate performance on two practically relevant localization tasks: loop closure detection and global localization. Our approach achieves major performance improvements over both existing and improved state-of-the-art systems.

Ming Xu, Niko Sünderhauf, Michael Milford

Visual localization techniques often comprise a hierarchical localization pipeline, with a visual place recognition module used as a coarse localizer to initialize a pose refinement stage. While improving the pose refinement step has been the focus of much recent research, most work on the coarse localization stage has focused on improvements like increased invariance to appearance change, without improving what can be loose error tolerances. In this letter, we propose two methods which adapt image retrieval techniques used for visual place recognition to the Bayesian state estimation formulation for localization. We demonstrate significant improvements to the localization accuracy of the coarse localization stage using our methods, whilst retaining state-of-the-art performance under severe appearance change. Using extensive experimentation on the Oxford RobotCar dataset, results show that our approach outperforms comparable state-of-the-art methods in terms of precision-recall performance for localizing image sequences. In addition, our proposed methods provides the flexibility to contextually scale localization latency in order to achieve these improvements. The improved initial localization estimate opens up the possibility of both improved overall localization performance and modified pose refinement techniques that leverage this improved spatial prior.

Dimity Miller, Niko Sünderhauf, Michael Milford et al.

Deployed into an open world, object detectors are prone to open-set errors, false positive detections of object classes not present in the training dataset. We propose GMM-Det, a real-time method for extracting epistemic uncertainty from object detectors to identify and reject open-set errors. GMM-Det trains the detector to produce a structured logit space that is modelled with class-specific Gaussian Mixture Models. At test time, open-set errors are identified by their low log-probability under all Gaussian Mixture Models. We test two common detector architectures, Faster R-CNN and RetinaNet, across three varied datasets spanning robotics and computer vision. Our results show that GMM-Det consistently outperforms existing uncertainty techniques for identifying and rejecting open-set detections, especially at the low-error-rate operating point required for safety-critical applications. GMM-Det maintains object detection performance, and introduces only minimal computational overhead. We also introduce a methodology for converting existing object detection datasets into specific open-set datasets to evaluate open-set performance in object detection.

Sourav Garg, Niko Sünderhauf, Feras Dayoub et al.

For robots to navigate and interact more richly with the world around them, they will likely require a deeper understanding of the world in which they operate. In robotics and related research fields, the study of understanding is often referred to as semantics, which dictates what does the world "mean" to a robot, and is strongly tied to the question of how to represent that meaning. With humans and robots increasingly operating in the same world, the prospects of human-robot interaction also bring semantics and ontology of natural language into the picture. Driven by need, as well as by enablers like increasing availability of training data and computational resources, semantics is a rapidly growing research area in robotics. The field has received significant attention in the research literature to date, but most reviews and surveys have focused on particular aspects of the topic: the technical research issues regarding its use in specific robotic topics like mapping or segmentation, or its relevance to one particular application domain like autonomous driving. A new treatment is therefore required, and is also timely because so much relevant research has occurred since many of the key surveys were published. This survey therefore provides an overarching snapshot of where semantics in robotics stands today. We establish a taxonomy for semantics research in or relevant to robotics, split into four broad categories of activity, in which semantics are extracted, used, or both. Within these broad categories we survey dozens of major topics including fundamentals from the computer vision field and key robotics research areas utilizing semantics, including mapping, navigation and interaction with the world. The survey also covers key practical considerations, including enablers like increased data availability and improved computational hardware, and major application areas where...

Quazi Marufur Rahman, Niko Sünderhauf, Feras Dayoub

During deployment, an object detector is expected to operate at a similar performance level reported on its testing dataset. However, when deployed onboard mobile robots that operate under varying and complex environmental conditions, the detector's performance can fluctuate and occasionally degrade severely without warning. Undetected, this can lead the robot to take unsafe and risky actions based on low-quality and unreliable object detections. We address this problem and introduce a cascaded neural network that monitors the performance of the object detector by predicting the quality of its mean average precision (mAP) on a sliding window of the input frames. The proposed cascaded network exploits the internal features from the deep neural network of the object detector. We evaluate our proposed approach using different combinations of autonomous driving datasets and object detectors.

Quazi Marufur Rahman, Niko Sünderhauf, Feras Dayoub

Performance monitoring of object detection is crucial for safety-critical applications such as autonomous vehicles that operate under varying and complex environmental conditions. Currently, object detectors are evaluated using summary metrics based on a single dataset that is assumed to be representative of all future deployment conditions. In practice, this assumption does not hold, and the performance fluctuates as a function of the deployment conditions. To address this issue, we propose an introspection approach to performance monitoring during deployment without the need for ground truth data. We do so by predicting when the per-frame mean average precision drops below a critical threshold using the detector's internal features. We quantitatively evaluate and demonstrate our method's ability to reduce risk by trading off making an incorrect decision by raising the alarm and absenting from detection.

David Hall, Ben Talbot, Suman Raj Bista et al.

Being able to explore an environment and understand the location and type of all objects therein is important for indoor robotic platforms that must interact closely with humans. However, it is difficult to evaluate progress in this area due to a lack of standardized testing which is limited due to the need for active robot agency and perfect object ground-truth. To help provide a standard for testing scene understanding systems, we present a new robot vision scene understanding challenge using simulation to enable repeatable experiments with active robot agency. We provide two challenging task types, three difficulty levels, five simulated environments and a new evaluation measure for evaluating 3D cuboid object maps. Our aim is to drive state-of-the-art research in scene understanding through enabling evaluation and comparison of active robotic vision systems.

Ben Talbot, David Hall, Haoyang Zhang et al.

We introduce BenchBot, a novel software suite for benchmarking the performance of robotics research across both photorealistic 3D simulations and real robot platforms. BenchBot provides a simple interface to the sensorimotor capabilities of a robot when solving robotics research problems; an interface that is consistent regardless of whether the target platform is simulated or a real robot. In this paper we outline the BenchBot system architecture, and explore the parallels between its user-centric design and an ideal research development process devoid of tangential robot engineering challenges. The paper describes the research benefits of using the BenchBot system, including: enhanced capacity to focus solely on research problems, direct quantitative feedback to inform research development, tools for deriving comprehensive performance characteristics, and submission formats which promote sharability and repeatability of research outcomes. BenchBot is publicly available (http://benchbot.org), and we encourage its use in the research community for comprehensively evaluating the simulated and real world performance of novel robotic algorithms.

Dimity Miller, Niko Sünderhauf, Michael Milford et al.

In open set recognition, deep neural networks encounter object classes that were unknown during training. Existing open set classifiers distinguish between known and unknown classes by measuring distance in a network's logit space, assuming that known classes cluster closer to the training data than unknown classes. However, this approach is applied post-hoc to networks trained with cross-entropy loss, which does not guarantee this clustering behaviour. To overcome this limitation, we introduce the Class Anchor Clustering (CAC) loss. CAC is a distance-based loss that explicitly trains known classes to form tight clusters around anchored class-dependent centres in the logit space. We show that training with CAC achieves state-of-the-art performance for distance-based open set classifiers on all six standard benchmark datasets, with a 15.2% AUROC increase on the challenging TinyImageNet, without sacrificing classification accuracy. We also show that our anchored class centres achieve higher open set performance than learnt class centres, particularly on object-based datasets and large numbers of training classes.

Krishan Rana, Vibhavari Dasagi, Ben Talbot et al.

Learning-based approaches often outperform hand-coded algorithmic solutions for many problems in robotics. However, learning long-horizon tasks on real robot hardware can be intractable, and transferring a learned policy from simulation to reality is still extremely challenging. We present a novel approach to model-free reinforcement learning that can leverage existing sub-optimal solutions as an algorithmic prior during training and deployment. During training, our gated fusion approach enables the prior to guide the initial stages of exploration, increasing sample-efficiency and enabling learning from sparse long-horizon reward signals. Importantly, the policy can learn to improve beyond the performance of the sub-optimal prior since the prior's influence is annealed gradually. During deployment, the policy's uncertainty provides a reliable strategy for transferring a simulation-trained policy to the real world by falling back to the prior controller in uncertain states. We show the efficacy of our Multiplicative Controller Fusion approach on the task of robot navigation and demonstrate safe transfer from simulation to the real world without any fine-tuning. The code for this project is made publicly available at https://sites.google.com/view/mcf-nav/home

Peter Corke, Feras Dayoub, David Hall et al.

The computer vision and robotics research communities are each strong. However progress in computer vision has become turbo-charged in recent years due to big data, GPU computing, novel learning algorithms and a very effective research methodology. By comparison, progress in robotics seems slower. It is true that robotics came later to exploring the potential of learning -- the advantages over the well-established body of knowledge in dynamics, kinematics, planning and control is still being debated, although reinforcement learning seems to offer real potential. However, the rapid development of computer vision compared to robotics cannot be only attributed to the former's adoption of deep learning. In this paper, we argue that the gains in computer vision are due to research methodology -- evaluation under strict constraints versus experiments; bold numbers versus videos.

Krishan Rana, Ben Talbot, Vibhavari Dasagi et al.

In this work we focus on improving the efficiency and generalisation of learned navigation strategies when transferred from its training environment to previously unseen ones. We present an extension of the residual reinforcement learning framework from the robotic manipulation literature and adapt it to the vast and unstructured environments that mobile robots can operate in. The concept is based on learning a residual control effect to add to a typical sub-optimal classical controller in order to close the performance gap, whilst guiding the exploration process during training for improved data efficiency. We exploit this tight coupling and propose a novel deployment strategy, switching Residual Reactive Navigation (sRRN), which yields efficient trajectories whilst probabilistically switching to a classical controller in cases of high policy uncertainty. Our approach achieves improved performance over end-to-end alternatives and can be incorporated as part of a complete navigation stack for cluttered indoor navigation tasks in the real world. The code and training environment for this project is made publicly available at https://sites.google.com/view/srrn/home.

Niko Sünderhauf

Emerging object-based SLAM algorithms can build a graph representation of an environment comprising nodes for robot poses and object landmarks. However, while this map will contain static objects such as furniture or appliances, many moveable objects (e.g. the car keys, the glasses, or a magazine), are not suitable as landmarks and will not be part of the map due to their non-static nature. We show that Graph Convolutional Networks can learn navigation policies to find such unmapped objects by learning to exploit the hidden probabilistic model that governs where these objects appear in the environment. The learned policies can generalise to object classes unseen during training by using word vectors that express semantic similarity as representations for object nodes in the graph. Furthermore, we show that the policies generalise to unseen environments with only minimal loss of performance. We demonstrate that pre-training the policy network with a proxy task can significantly speed up learning, improving sample efficiency.

John Skinner, David Hall, Haoyang Zhang et al.

We introduce a new challenge for computer and robotic vision, the first ACRV Robotic Vision Challenge, Probabilistic Object Detection. Probabilistic object detection is a new variation on traditional object detection tasks, requiring estimates of spatial and semantic uncertainty. We extend the traditional bounding box format of object detection to express spatial uncertainty using gaussian distributions for the box corners. The challenge introduces a new test dataset of video sequences, which are designed to more closely resemble the kind of data available to a robotic system. We evaluate probabilistic detections using a new probability-based detection quality (PDQ) measure. The goal in creating this challenge is to draw the computer and robotic vision communities together, toward applying object detection solutions for practical robotics applications.

Quazi Marufur Rahman, Niko Sünderhauf, Feras Dayoub

Object detection is an integral part of an autonomous vehicle for its safety-critical and navigational purposes. Traffic signs as objects play a vital role in guiding such systems. However, if the vehicle fails to locate any critical sign, it might make a catastrophic failure. In this paper, we propose an approach to identify traffic signs that have been mistakenly discarded by the object detector. The proposed method raises an alarm when it discovers a failure by the object detector to detect a traffic sign. This approach can be useful to evaluate the performance of the detector during the deployment phase. We trained a single shot multi-box object detector to detect traffic signs and used its internal features to train a separate false negative detector (FND). During deployment, FND decides whether the traffic sign detector (TSD) has missed a sign or not. We are using precision and recall to measure the accuracy of FND in two different datasets. For 80% recall, FND has achieved 89.9% precision in Belgium Traffic Sign Detection dataset and 90.8% precision in German Traffic Sign Recognition Benchmark dataset respectively. To the best of our knowledge, our method is the first to tackle this critical aspect of false negative detection in robotic vision. Such a fail-safe mechanism for object detection can improve the engagement of robotic vision systems in our daily life.

David Hall, Feras Dayoub, John Skinner et al.

We introduce Probabilistic Object Detection, the task of detecting objects in images and accurately quantifying the spatial and semantic uncertainties of the detections. Given the lack of methods capable of assessing such probabilistic object detections, we present the new Probability-based Detection Quality measure (PDQ).Unlike AP-based measures, PDQ has no arbitrary thresholds and rewards spatial and label quality, and foreground/background separation quality while explicitly penalising false positive and false negative detections. We contrast PDQ with existing mAP and moLRP measures by evaluating state-of-the-art detectors and a Bayesian object detector based on Monte Carlo Dropout. Our experiments indicate that conventional object detectors tend to be spatially overconfident and thus perform poorly on the task of probabilistic object detection. Our paper aims to encourage the development of new object detection approaches that provide detections with accurately estimated spatial and label uncertainties and are of critical importance for deployment on robots and embodied AI systems in the real world.

Vibhavari Dasagi, Robert Lee, Serena Mou et al.

Current end-to-end deep Reinforcement Learning (RL) approaches require jointly learning perception, decision-making and low-level control from very sparse reward signals and high-dimensional inputs, with little capability of incorporating prior knowledge. This results in prohibitively long training times for use on real-world robotic tasks. Existing algorithms capable of extracting task-level representations from high-dimensional inputs, e.g. object detection, often produce outputs of varying lengths, restricting their use in RL methods due to the need for neural networks to have fixed length inputs. In this work, we propose a framework that combines deep sets encoding, which allows for variable-length abstract representations, with modular RL that utilizes these representations, decoupling high-level decision making from low-level control. We successfully demonstrate our approach on the robot manipulation task of object sorting, showing that this method can learn effective policies within mere minutes of highly simplified simulation. The learned policies can be directly deployed on a robot without further training, and generalize to variations of the task unseen during training.

Natalie Jablonsky, Michael Milford, Niko Sünderhauf

Current approaches to object-oriented SLAM lack the ability to incorporate prior knowledge of the scene geometry, such as the expected global orientation of objects. We overcome this limitation by proposing a geometric factor that constrains the global orientation of objects in the map, depending on the objects' semantics. This new geometric factor is a first example of how semantics can inform and improve geometry in object-oriented SLAM. We implement the geometric factor for the recently proposed QuadricSLAM that represents landmarks as dual quadrics. The factor probabilistically models the quadrics' major axes to be either perpendicular to or aligned with the direction of gravity, depending on their semantic class. Our experiments on simulated and real-world datasets show that using the proposed factors to incorporate prior knowledge improves both the trajectory and landmark quality.

Dimity Miller, Feras Dayoub, Michael Milford et al.

There has been a recent emergence of sampling-based techniques for estimating epistemic uncertainty in deep neural networks. While these methods can be applied to classification or semantic segmentation tasks by simply averaging samples, this is not the case for object detection, where detection sample bounding boxes must be accurately associated and merged. A weak merging strategy can significantly degrade the performance of the detector and yield an unreliable uncertainty measure. This paper provides the first in-depth investigation of the effect of different association and merging strategies. We compare different combinations of three spatial and two semantic affinity measures with four clustering methods for MC Dropout with a Single Shot Multi-Box Detector. Our results show that the correct choice of affinity-clustering combination can greatly improve the effectiveness of the classification and spatial uncertainty estimation and the resulting object detection performance. We base our evaluation on a new mix of datasets that emulate near open-set conditions (semantically similar unknown classes), distant open-set conditions (semantically dissimilar unknown classes) and the common closed-set conditions (only known classes).

Jake Bruce, Niko Sünderhauf, Piotr Mirowski et al.

Model-free reinforcement learning has recently been shown to be effective at learning navigation policies from complex image input. However, these algorithms tend to require large amounts of interaction with the environment, which can be prohibitively costly to obtain on robots in the real world. We present an approach for efficiently learning goal-directed navigation policies on a mobile robot, from only a single coverage traversal of recorded data. The navigation agent learns an effective policy over a diverse action space in a large heterogeneous environment consisting of more than 2km of travel, through buildings and outdoor regions that collectively exhibit large variations in visual appearance, self-similarity, and connectivity. We compare pretrained visual encoders that enable precomputation of visual embeddings to achieve a throughput of tens of thousands of transitions per second at training time on a commodity desktop computer, allowing agents to learn from millions of trajectories of experience in a matter of hours. We propose multiple forms of computationally efficient stochastic augmentation to enable the learned policy to generalise beyond these precomputed embeddings, and demonstrate successful deployment of the learned policy on the real robot without fine tuning, despite environmental appearance differences at test time. The dataset and code required to reproduce these results and apply the technique to other datasets and robots is made publicly available at rl-navigation.github.io/deployable.

Niko Sünderhauf, Oliver Brock, Walter Scheirer et al.

The application of deep learning in robotics leads to very specific problems and research questions that are typically not addressed by the computer vision and machine learning communities. In this paper we discuss a number of robotics-specific learning, reasoning, and embodiment challenges for deep learning. We explain the need for better evaluation metrics, highlight the importance and unique challenges for deep robotic learning in simulation, and explore the spectrum between purely data-driven and model-driven approaches. We hope this paper provides a motivating overview of important research directions to overcome the current limitations, and help fulfill the promising potentials of deep learning in robotics.

Lachlan Nicholson, Michael Milford, Niko Sünderhauf

In this paper, we use 2D object detections from multiple views to simultaneously estimate a 3D quadric surface for each object and localize the camera position. We derive a SLAM formulation that uses dual quadrics as 3D landmark representations, exploiting their ability to compactly represent the size, position and orientation of an object, and show how 2D object detections can directly constrain the quadric parameters via a novel geometric error formulation. We develop a sensor model for object detectors that addresses the challenge of partially visible objects, and demonstrate how to jointly estimate the camera pose and constrained dual quadric parameters in factor graph based SLAM with a general perspective camera.

Peter Anderson, Qi Wu, Damien Teney et al.

A robot that can carry out a natural-language instruction has been a dream since before the Jetsons cartoon series imagined a life of leisure mediated by a fleet of attentive robot helpers. It is a dream that remains stubbornly distant. However, recent advances in vision and language methods have made incredible progress in closely related areas. This is significant because a robot interpreting a natural-language navigation instruction on the basis of what it sees is carrying out a vision and language process that is similar to Visual Question Answering. Both tasks can be interpreted as visually grounded sequence-to-sequence translation problems, and many of the same methods are applicable. To enable and encourage the application of vision and language methods to the problem of interpreting visually-grounded navigation instructions, we present the Matterport3D Simulator -- a large-scale reinforcement learning environment based on real imagery. Using this simulator, which can in future support a range of embodied vision and language tasks, we provide the first benchmark dataset for visually-grounded natural language navigation in real buildings -- the Room-to-Room (R2R) dataset.

Dimity Miller, Lachlan Nicholson, Feras Dayoub et al.

Dropout Variational Inference, or Dropout Sampling, has been recently proposed as an approximation technique for Bayesian Deep Learning and evaluated for image classification and regression tasks. This paper investigates the utility of Dropout Sampling for object detection for the first time. We demonstrate how label uncertainty can be extracted from a state-of-the-art object detection system via Dropout Sampling. We evaluate this approach on a large synthetic dataset of 30,000 images, and a real-world dataset captured by a mobile robot in a versatile campus environment. We show that this uncertainty can be utilized to increase object detection performance under the open-set conditions that are typically encountered in robotic vision. A Dropout Sampling network is shown to achieve a 12.3% increase in recall (for the same precision score as a standard network) and a 15.1% increase in precision (for the same recall score as the standard network).

Trung Pham, Thanh-Toan Do, Niko Sünderhauf et al.

This paper presents SceneCut, a novel approach to jointly discover previously unseen objects and non-object surfaces using a single RGB-D image. SceneCut's joint reasoning over scene semantics and geometry allows a robot to detect and segment object instances in complex scenes where modern deep learning-based methods either fail to separate object instances, or fail to detect objects that were not seen during training. SceneCut automatically decomposes a scene into meaningful regions which either represent objects or scene surfaces. The decomposition is qualified by an unified energy function over objectness and geometric fitting. We show how this energy function can be optimized efficiently by utilizing hierarchical segmentation trees. Moreover, we leverage a pre-trained convolutional oriented boundary network to predict accurate boundaries from images, which are used to construct high-quality region hierarchies. We evaluate SceneCut on several different indoor environments, and the results show that SceneCut significantly outperforms all the existing methods.

Niko Sünderhauf, Michael Milford

Research in Simultaneous Localization And Mapping (SLAM) is increasingly moving towards richer world representations involving objects and high level features that enable a semantic model of the world for robots, potentially leading to a more meaningful set of robot-world interactions. Many of these advances are grounded in state-of-the-art computer vision techniques primarily developed in the context of image-based benchmark datasets, leaving several challenges to be addressed in adapting them for use in robotics. In this paper, we derive a formulation for Simultaneous Localization And Mapping (SLAM) that uses dual quadrics as 3D landmark representations, and show how 2D bounding boxes (such as those typically obtained from visual object detection systems) can directly constrain the quadric parameters. Our paper demonstrates how to jointly estimate the robot pose and dual quadric parameters in factor graph based SLAM with a general perspective camera, and covers the use-cases of a robot moving with a monocular camera with and without the availability of additional depth information.

Sean McMahon, Niko Sünderhauf, Ben Upcroft et al.

Trip hazards are a significant contributor to accidents on construction and manufacturing sites, where over a third of Australian workplace injuries occur [1]. Current safety inspections are labour intensive and limited by human fallibility,making automation of trip hazard detection appealing from both a safety and economic perspective. Trip hazards present an interesting challenge to modern learning techniques because they are defined as much by affordance as by object type; for example wires on a table are not a trip hazard, but can be if lying on the ground. To address these challenges, we conduct a comprehensive investigation into the performance characteristics of 11 different colour and depth fusion approaches, including 4 fusion and one non fusion approach; using colour and two types of depth images. Trained and tested on over 600 labelled trip hazards over 4 floors and 2000m$\mathrm{^{2}}$ in an active construction site,this approach was able to differentiate between identical objects in different physical configurations (see Figure 1). Outperforming a colour-only detector, our multi-modal trip detector fuses colour and depth information to achieve a 4% absolute improvement in F1-score. These investigative results and the extensive publicly available dataset moves us one step closer to assistive or fully automated safety inspection systems on construction sites.

Feras Dayoub, Niko Sünderhauf, Peter Corke

We investigate different strategies for active learning with Bayesian deep neural networks. We focus our analysis on scenarios where new, unlabeled data is obtained episodically, such as commonly encountered in mobile robotics applications. An evaluation of different strategies for acquisition, updating, and final training on the CIFAR-10 dataset shows that incremental network updates with final training on the accumulated acquisition set are essential for best performance, while limiting the amount of required human labeling labor.

Adam Tow, Niko Sünderhauf, Sareh Shirazi et al.

We propose to learn tasks directly from visual demonstrations by learning to predict the outcome of human and robot actions on an environment. We enable a robot to physically perform a human demonstrated task without knowledge of the thought processes or actions of the human, only their visually observable state transitions. We evaluate our approach on two table-top, object manipulation tasks and demonstrate generalisation to previously unseen states. Our approach reduces the priors required to implement a robot task learning system compared with the existing approaches of Learning from Demonstration, Reinforcement Learning and Inverse Reinforcement Learning.

Niko Sünderhauf, Trung T. Pham, Yasir Latif et al.

For intelligent robots to interact in meaningful ways with their environment, they must understand both the geometric and semantic properties of the scene surrounding them. The majority of research to date has addressed these mapping challenges separately, focusing on either geometric or semantic mapping. In this paper we address the problem of building environmental maps that include both semantically meaningful, object-level entities and point- or mesh-based geometrical representations. We simultaneously build geometric point cloud models of previously unseen instances of known object classes and create a map that contains these object models as central entities. Our system leverages sparse, feature-based RGB-D SLAM, image-based deep-learning object detection and 3D unsupervised segmentation.

Niko Sünderhauf, Feras Dayoub, Sean McMahon et al.

In this paper we focus on the challenging problem of place categorization and semantic mapping on a robot without environment-specific training. Motivated by their ongoing success in various visual recognition tasks, we build our system upon a state-of-the-art convolutional network. We overcome its closed-set limitations by complementing the network with a series of one-vs-all classifiers that can learn to recognize new semantic classes online. Prior domain knowledge is incorporated by embedding the classification system into a Bayesian filter framework that also ensures temporal coherence. We evaluate the classification accuracy of the system on a robot that maps a variety of places on our campus in real-time. We show how semantic information can boost robotic object detection performance and how the semantic map can be used to modulate the robot's behaviour during navigation tasks. The system is made available to the community as a ROS module.