Soft Actor-Critic (SAC) has achieved notable success in continuous control tasks but struggles in sparse reward settings, where infrequent rewards make efficient exploration challenging. While novelty-based exploration methods address this issue by encouraging the agent to explore novel states, they are not trivial to apply to SAC. In particular, managing the interaction between novelty-based exploration and SAC’s stochastic policy can lead to inefficient exploration and redundant sample collection. In this paper, we propose KEA (Keeping Exploration Alive) which tackles the inefficiencies in balancing exploration strategies when combining SAC with novelty-based exploration. KEA integrates a novelty-augmented SAC with a standard SAC agent, proactively coordinated via a switching mechanism. This coordination allows the agent to maintain stochasticity in high-novelty regions, enhancing exploration efficiency and reducing repeated sample collection. We first analyze this potential issue in a 2D navigation task, and then evaluate KEA on the DeepSea hard-exploration benchmark as well as sparse reward control tasks from the DeepMind Control Suite. Compared to state-of-the-art novelty-based exploration baselines, our experiments show that KEA significantly improves learning efficiency and robustness in sparse reward setups.

Many practically relevant robot grasping problems feature a target object for which all grasps are occluded, e.g., by the environment. Single-shot grasp planning invariably fails in such scenarios. Instead, it is necessary to first manipulate the object into a configuration that affords a grasp. We solve this problem by learning a sequence of actions that utilize the environment to change the object’s pose. Concretely, we employ hierarchical reinforcement learning to combine a sequence of learned parameterized manipulation primitives. By learning the low-level manipulation policies, our approach can control the object’s state through exploiting interactions between the object, the gripper, and the environment. Designing such a complex behavior analytically would be infeasible under uncontrolled conditions, as an analytic approach requires accurate physical modeling of the interaction and contact dynamics. In contrast, we learn a hierarchical policy model that operates directly on depth perception data, without the need for object detection, pose estimation, or manual design of controllers. We evaluate our approach on picking box-shaped objects of various weight, shape, and friction properties from a constrained table-top workspace. Our method transfers to a real robot and is able to successfully complete the object picking task in 98% of experimental trials.

Abstract One-step reinforcement learning explanation methods account for individual actions but fail to consider the agent’s future behavior, which can make their interpretation ambiguous. We propose to address this limitation by providing hierarchical goals as context for one-step explanations. By considering the current hierarchical goal as a context, one-step explanations can be interpreted with higher certainty, as the agent’s future behavior is more predictable. We combine reward decomposition with hierarchical reinforcement learning into a novel explainable reinforcement learning framework, which yields more interpretable, goal-contextualized one-step explanations. With a qualitative analysis of one-step reward decomposition explanations, we first show that their interpretability is indeed limited in scenarios with multiple, different optimal policies—a characteristic shared by other one-step explanation methods. Then, we show that our framework retains high interpretability in such cases, as the hierarchical goal can be considered as context for the explanation. To the best of our knowledge, our work is the first to investigate hierarchical goals not as an explanation directly but as additional context for one-step reinforcement learning explanations.

Conventional methods to autonomous grasping rely on a pre-computed database with known objects to synthesize grasps, which is not possible for novel objects. On the other hand, recently proposed deep learning-based approaches have demonstrated the ability to generalize grasp for unknown objects. However, grasp generation still remains a challenging problem, especially in cluttered environments under partial occlusion. In this work, we propose an end-to-end deep learning approach for generating 6-DOF collision-free grasps given a 3D scene point cloud. To build robustness to occlusion, the proposed model generates candidates by casting votes and accumulating evidence for feasible grasp configurations. We exploit contextual information by encoding the dependency of objects in the scene into features to boost the performance of grasp generation. The contextual information enables our model to increase the likelihood that the generated grasps are collision-free. Our experimental results confirm that the proposed system performs favorably in terms of predicting object grasps in cluttered environments in comparison to the current state of the art methods.

Traditional approaches to manipulating the state of deformable linear objects (DLOs) - i.e., cables, ropes - rely on model-based planning. However, constructing an accurate dynamic model of a DLO is challenging due to the complexity of interactions and a high number of degrees of freedom. This renders the task of achieving a desired DLO shape particularly difficult and motivates the use of model-free alternatives, which while maintaining generality suffer from a high sample complexity. In this paper, we bridge the gap between these fundamentally different approaches and propose a framework that learns dynamic models of DLOs through trial-and-error interaction. Akin to model-based reinforcement learning (RL), we interleave learning and exploration to solve a 3D shape control task for a DLO. Our approach requires only a fraction of the interaction samples of the current state-of-the-art model-free RL alternatives to achieve superior shape control performance. Unlike offline model learning, our approach does not require expert knowledge for data collection, retains the ability to explore, and automatically selects relevant experience.

Modeling dynamics of deformable linear objects (DLOs), such as cables, hoses, sutures, and catheters, is an important and challenging problem for many robotic manipulation applications. In this paper, we propose the first method to model and learn full 3D dynamics of DLOs from data. Our approach is capable of capturing the complex twisting and bending dynamics of DLOs and allows local effects to propagate globally. To this end, we adapt the interaction network (IN) dynamics learning method for capturing the interaction between neighboring segments in a DLO and augment it with a recurrent model for propagating interaction effects along the length of a DLO. For learning twisting and bending dynamics in 3D, we also introduce a new suitable representation of DLO segments and their relationships. Unlike the original IN method, our model learns to propagate the effects of local interaction between neighboring segments to each segment in the chain within a single time step, without the need for iterated propagation steps. Evaluation of our model with synthetic and newly collected real-world data shows better accuracy and generalization in short-term and long- term predictions than the current state of the art. We further integrate our learned model in a model predictive control scheme and use it to successfully control the shape of a DLO. Our implementation is available at https://gitsvn-nt.oru.se/ammlab–public/in–bilstm.

Creating maps is an essential task in robotics and provides the basis for effective planning and navigation. In this paper, we learn a compact and continuous implicit surface map of an environment from a stream of range data with known poses. For this, we create and incrementally adjust an ensemble of approximate Gaussian process (GP) experts which are each responsible for a different part of the map. Instead of inserting all arriving data into the GP models, we greedily trade-off between model complexity and prediction error. Our algorithm therefore uses less resources on areas with few geometric features and more where the environment is rich in variety. We evaluate our approach on synthetic and real-world data sets and analyze sensitivity to parameters and measurement noise. The results show that we can learn compact and accurate implicit surface models under different conditions, with a performance comparable to or better than that of exact GP regression with subsampled data.

Virtual reality (VR) is regarded as a useful tool for teleoperation systems and provides operators with immersive visual feedback on the robot and the environment. However, without any haptic feedback or physical constructions, VR-based teleoperation systems normally suffer from poor maneuverability, and operational faults may be caused in some fine movements. In this article, we employ mixed reality (MR), which combines real and virtual worlds, to develop a novel teleoperation system. A new system design and control algorithms are proposed. For the system design, an MR interface is developed based on a virtual environment augmented with real-time data from the task space with the goal of enhancing the operator's visual perception. To allow the operator to be freely decoupled from the control loop and offload the operator's burden, a new interaction proxy is proposed to control the robot. For the control algorithms, two control modes are introduced to improve the long-distance movements and fine movements of the MR-based teleoperation system. In addition, a set of fuzzy-logic-based methods are proposed to regulate the orientation, position, velocity, and force of the robot to enhance the system's maneuverability and address potential operational faults. A barrier Lyapunov function and a backstepping method are leveraged to design the control laws and simultaneously guarantee the system's stability under state constraints. Experiments conducted using a six-degree-of-freedom robotic arm prove the feasibility of the system.

Voxel volumes are simple to implement and lend themselves to many of the tools and algorithms available for 2D images. However, the additional dimension of voxels may be costly to manage in memory when mapping large spaces at high resolutions. While lowering the resolution and using interpolation is common work-around, in the literature we often find that authors either use trilinear interpolation or nearest neighbors and rarely any of the intermediate options. This paper presents a survey of geometric interpolation methods for voxel-based map representations. In particular we study the truncated signed distance field (TSDF) and the impact of using fewer than 8 samples to perform interpolation within a depth-camera pose tracking and mapping scenario. We find that lowering the number of samples fetched to perform the interpolation results in performance similar to the commonly used trilinear interpolation method, but leads to higher frame-rates. We also report that lower bit-depth generally leads to performance degradation, though not as much as may be expected, with voxels containing as few as 3 bits sometimes resulting in adequate estimation of camera trajectories.

This article presents an approach for assisted teleoperation of a robot arm, formulated within a real-time stack-of-tasks (SoT)whole-body motion control framework. The approach leverages the hierarchical nature of the SoT framework to integrate operator commands with assistive tasks, such as joint limit and obstacle avoidance or automatic gripper alignment. Thereby some aspects of the teleoperation problem are delegated to the controller and carried out autonomously. The key contributions of this work are two-fold: the first is a method for unobtrusive integration of autonomy in a telemanip-ulation system; and the second is a user study evaluation of the proposed system in the context of teleoperated pick-and-place tasks. The proposed approach of assistive control was found to result in higher grasp success rates and shorter trajectories than achieved through manual control, without incurring additional cognitive load to the operator.

Local scan registration approaches commonly only utilize ego-motion estimates (e.g. odometry) as an initial pose guess in an iterative alignment procedure. This paper describes a new method to incorporate ego-motion estimates, including uncertainty, into the objective function of a registration algorithm. The proposed approach is particularly suited for feature-poor and self-similar environments, which typically present challenges to current state of the art registration algorithms. Experimental evaluation shows significant improvements in accuracy when using data acquired by Automatic Guided Vehicles (AGVs) in industrial production and warehouse environments.

Grasping systems that build upon meticulously planned hand postures rely on precise knowledge of object geometry, mass and frictional properties — assumptions which are often violated in practice. In this work, we propose an alternative solution to the problem of grasp acquisition in simple autonomous pick and place scenarios, by utilizing the concept of grasp envelopes: sets of constraints on gripper postures. We propose a fast method for extracting grasp envelopes for objects that fit within a known shape category, placed in an unknown environment. Our approach is based on grasp envelope primitives, which encode knowledge of human grasping strategies. We use environment models, reconstructed from noisy sensor observations, to refine the grasp envelope primitives and extract bounded envelopes of collision-free gripper postures. Also, we evaluate the envelope extraction procedure both in a stand alone fashion, as well as an integrated component of an autonomous picking system.

Given that 3D scan matching is such a central part of the perception pipeline for robots, thorough and large-scale investigations of scan matching performance are still surprisingly few. A crucial part of the scientific method is to perform experiments that can be replicated by other researchers in order to compare different results. In light of this fact, this paper presents a thorough comparison of 3D scan registration algorithms using a recently published benchmark protocol which makes use of a publicly available challenging data set that covers a wide range of environments. In particular, we evaluate two types of recent 3D registration algorithms - one local and one global. Both approaches take local surface structure into account, rather than matching individual points. After well over 100 000 individual tests, we conclude that algorithms using the normal distributions transform (NDT) provides accurate results compared to a modern implementation of the iterative closest point (ICP) method, when faced with scan data that has little overlap and weak geometric structure. We also demonstrate that the minimally uncertain maximum consensus (MUMC) algorithm provides accurate results in structured environments without needing an initial guess, and that it provides useful measures to detect whether it has succeeded or not. We also propose two amendments to the experimental protocol, in order to provide more valuable results in future implementations.

Autonomous navigation in real-world industrial environments is a challenging task in many respects. One of the key open challenges is fast planning and execution of trajectories to reach arbitrary target positions and orientations with high accuracy and precision, while taking into account non-holonomic vehicle constraints. In recent years, lattice-based motion planners have been successfully used to generate kinematically and kinodynamically feasible motions for non-holonomic vehicles. However, the discretized nature of these algorithms induces discontinuities in both state and control space of the obtained trajectories, resulting in a mismatch between the achieved and the target end pose of the vehicle. As endpose accuracy is critical for the successful loading and unloading of cargo in typical industrial applications, automatically planned paths have not be widely adopted in commercial AGV systems. The main contribution of this paper addresses this shortcoming by introducing a path smoothing approach, which builds on the output of a lattice-based motion planner to generate smooth drivable trajectories for non-holonomic industrial vehicles. In real world tests presented in this paper we demonstrate that the proposed approach is fast enough for online use (it computes trajectories faster than they can be driven) and highly accurate. In 100 repetitions we achieve mean end-point pose errors below 0.01 meters in translation and 0.002 radians in orientation. Even the maximum errors are very small: only 0.02 meters in translation and 0.008 radians in orientation.

In this work we tackle the problem of planning grasps for an underactuated gripper which enable it to retrieve target objects from a cluttered environment. Furthermore, we investigate how additional manipulation capabilities of the gripping device, provided by active surfaces on the inside of the fingers, can lead to performance improvement in the grasp execution process. To this end, we employ a simple strategy, in which the target object is `pulled-in' towards the palm during grasping which results in firm enveloping grasps. We show the effectiveness of the suggested methods by means of experiments conducted in a real-world scenario.

In this paper we present an inspection robot to produce gas distribution maps and localize gas sources in large outdoor environments. The robot is equipped with a 3D laser range finder and a remote gas sensor that returns integral concentration measurements. We apply principles of tomography to create a spatial gas distribution model from integral gas concentration measurements. The gas distribution algorithm is framed as a convex optimization problem and it models the mean distribution and the fluctuations of gases. This is important since gas dispersion is not an static phenomenon and furthermore, areas of high fluctuation can be correlated with the location of an emitting source. We use a compact surface representation created from the measurements of the 3D laser range finder with a state of the art mapping algorithm to get a very accurate localization and estimation of the path of the laser beams. In addition, a conic model for the beam of the remote gas sensor is introduced. We observe a substantial improvement in the gas source localization capabilities over previous state-of-the-art in our evaluation carried out in an open field environment.

This paper presents a method for improving the quality of distorted 3D point clouds made from a vehicle equipped with a laser scanner moving over uneven terrain. Existing methods that use 3D point-cloud data (for tasks such as mapping, localisation, and object detection) typically assume that each point cloud is accurate. For autonomous robots moving in rough terrain, it is often the case that the vehicle moves a substantial amount during the acquisition of one point cloud, in which case the data will be distorted. The method proposed in this paper is capable of increasing the accuracy of 3D point clouds, without assuming any specific features of the environment (such as planar walls), without resorting to a “stop-scan-go” approach, and without relying on specialised and expensive hardware. Each new point cloud is matched to the previous using normal-distribution-transform (NDT) registration, after which a mini-loop closure is performed with a local, per-scan, graph-based SLAM method. The proposed method increases the accuracy of both the measured platform trajectory and the point cloud. The method is validated on both real-world and simulated data.

Autonomous vehicles operating in real-world industrial environments have to overcome numerous challenges, chief among which is the creation and maintenance of consistent 3D world models. This paper proposes to address the challenges of online real-world mapping by building upon previous work on compact spatial representation and formulating a novel 3D mapping approach - the Normal Distributions Transform Occupancy Map (NDT-OM). The presented algorithm enables accurate real-time 3D mapping in large-scale dynamic environments employing a recursive update strategy. In addition, the proposed approach can seamlessly provide maps at multiple resolutions allowing for fast utilization in high-level functions such as localization or path planning. Compared to previous approaches that use the NDT representation, the proposed NDT-OM formulates an exact and efficient recursive update formulation and models the full occupancy of the map.

In this work we propose a method to effectively remove noise from depth images obtained with a commodity structured light sensor. The proposed approach fuses data into a consistent frame of reference over time, thus utilizing prior depth measurements and viewpoint information in the noise removal process. The effectiveness of the approach is compared to two state of the art, single-frame denoising methods in the context of feature descriptor matching and keypoint detection stability. To make more general statements about the effect of noise removal in these applications, we extend a method for evaluating local image gradient feature descriptors to the domain of 3D shape descriptors. We perform a comparative study of three classes of such descriptors: Normal Aligned Radial Features, Fast Point Feature Histograms and Depth Kernel Descriptors; and evaluate their performance on a real-world industrial application data set. We demonstrate that noise removal enabled by the dense map representation results in major improvements in matching across all classes of descriptors as well as having a substantial positive impact on keypoint detection reliability.

Ego-motion estimation and environment mapping are two recurring problems in the field of robotics. In this work we propose a simple on-line method for tracking the pose of a depth camera in six degrees of freedom and simultaneously maintaining an updated 3D map, represented as a truncated signed distance function. The distance function representation implicitly encodes surfaces in 3D-space and is used directly to define a cost function for accurate registration of new data. The proposed algorithm is highly parallel and achieves good accuracy compared to state of the art methods. It is suitable for reconstructing single household items, workspace environments and small rooms at near real-time rates, making it practical for use on modern CPU hardware.

Industrial applications often impose hard requirements on the precision of autonomous vehicle systems. As a consequence industrial Automatically Guided Vehicle (AGV) systems still use high-cost infrastructure based positioning solutions. In this paper we propose a map based localization method that fulfills the requirements on precision and repeatability, typical for industrial application scenarios. The proposed method - Normal Distributions Transform Monte Carlo Localization (NDT-MCL) is based on a well established probabilistic framework. In a novel contribution, we formulate the MCL localization approach using the Normal Distributions Transform (NDT) as an underlying representation for both map and sensor data. By relaxing the hard discretization assumption imposed by grid-map models and utilizing the piece-wise continuous NDT representation the proposed algorithm achieves substantially improved accuracy and repeatability. The proposed NDT-MCL algorithm is evaluated using offline data sets from both a laboratory and a real-world industrial environments. Additionally, we report a comparison of the proposed algorithm to grid-based MCL and to a commercial localization system when used in a closed-loop with the control system of an AGV platform. In all tests the proposed algorithm is demonstrated to provide performance superior to that of standard grid-based MCL and comparable to the performance of the commercial infrastructure based positioning system.

Autonomous vehicles operating in real-world industrial environments have to overcome numerous challenges, chief among which is the creation and maintenance of consistent 3D world models. This paper focuses on a particularly important challenge: mapping in dynamic environments. We introduce several improvements to the recently proposed Normal Distributions Transform Occupancy Map (NDT-OM) aimed for efficient mapping in dynamic environments. A careful consistency analysis is given based on convergence and similarity metrics specifically designed for evaluation of NDT maps in dynamic environments. We show that in the context of mapping with known poses the proposed method results in improved consistency and in superior runtime performance, when compared against 3D occupancy grids at the same size and resolution. Additionally, we demonstrate that NDT-OM features real-time performance in a highly dynamic 3D mapping and tracking scenario with centimeter accuracy over a 1.5km trajectory.

Autonomous vehicles operating in real-world industrial environments have to overcome numerous challenges, chief among which are the creation of consistent 3D world models and the simultaneous tracking of the vehicle pose with respect to the created maps. In this paper we integrate two recently proposed algorithms in an online, near-realtime mapping and tracking system. Using the Normal Distributions Transform (NDT), a sparse Gaussian Mixture Model, for representation of 3D range scan data, we propose a frame-to-model registration and data fusion algorithm - NDT Fusion. The proposed approach uses a submap indexing system to achieve operation in arbitrarily-sized environments. The approach is evaluated on a publicly available city-block sized data set, achieving accuracy and runtime performance significantly better than current state of the art. In addition, the system is evaluated on a data set covering ten hours of operation and a trajectory of 7.2km in a real-world industrial environment, achieving centimeter accuracy at update rates of 5-10 Hz.

The synthesis and evaluation of multi-fingered grasps on complex objects is a challenging problem that has received much attention in the robotics community. Although several promising approaches have been developed, applications to real-world systems are limited to simple objects or gripper configurations. The paradigm of Independent Contact Regions (ICRs) has been proposed as a way to increase the tolerance to grasp positioning errors. This concept is well established, though only on precise geometric object models. This work is concerned with the application of the ICR paradigm to models reconstructed from real-world range data. We propose a method for increasing the robustness of grasp synthesis on uncertain geometric models. The sensitivity of the ICR algorithm to noisy data is evaluated and a filtering approach is proposed to improve the quality of the final result.

Point set registration-the task of finding the best fitting alignment between two sets of point samples, is an important problem in mobile robotics. This article proposes a novel registration algorithm, based on the distance between Three-Dimensional Normal Distributions Transforms. 3D-NDT models - a sub-class of Gaussian Mixture Models with uniformly weighted, largely disjoint components, can be quickly computed from range point data. The proposed algorithm constructs 3D-NDT representations of the input point sets and then formulates an objective function based on the L2distance between the considered models. Analytic first and second order derivatives of the objective function are computed and used in a standard Newton method optimization scheme, to obtain the best-fitting transformation. The proposed algorithm is evaluated and shown to be more accurate and faster, compared to a state of the art implementation of the Iterative Closest Point and 3D-NDT Point-to-Distribution algorithms.

The Three-Dimensional Normal Distributions Transform (3D-NDT) is a spatial modeling technique with applications in point set registration, scan similarity comparison, change detection and path planning. This work concentrates on evaluating three common variations of the 3D-NDT in terms of accuracy of representing sampled semi-structured environments. In a novel approach to spatial representation quality measurement, the 3D geometrical modeling task is formulated as a classification problem and its accuracy is evaluated with standard machine learning performance metrics. In this manner the accuracy of the 3D-NDT variations is shown to be comparable to, and in some cases to outperform that of the standard occupancy grid mapping model.

Planning feasible paths in fully three-dimensional environments is a challenging problem. Application of existing algorithms typically requires the use of limited 3D representations that discard potentially useful information. This article proposes a novel approach to path planning that utilizes a full 3D representation directly: the Three-Dimensional Normal Distributions Transform (3D-NDT). The well known wavefront planner is modified to use 3D-NDT as a basis for map representation and evaluated using both indoor and outdoor data sets. The use of 3D-NDT for path planning is thus demonstrated to be a viable choice with good expressive capabilities.

Exploration of unknown environments remains one of the fundamental problems of mobile robotics. It is also a prime example for a task that can benefit significantly from multi-robot teams. We present an integrated system for semi-autonomous cooperative exploration, augmented by an intuitive user interface for efficient human supervision and control. In this preliminary study we demonstrate the effectiveness of the system as a whole and the intuitive interface in particular. Congruent with previous findings, results confirm that having a human in the loop improves task performance, especially with larger numbers of robots. Specific to our interface, we find that even untrained operators can efficiently manage a decently sized team of robots.