IROS 2023 论文列表

For a robot to repair its own error, it must first know it has made a mistake. One way that people detect errors is from the implicit reactions from bystanders - their confusion, smirks, or giggles clue us in that something unexpected occurred. To enable robots to detect and act on bystander responses to task failures, we developed a novel method to elicit bystander responses to human and robot errors. Using 46 different stimulus videos featuring a variety of human and machine task failures, we collected a total of 2,452 webcam videos of human reactions from 54 participants. To test the viability of the collected data, we used the bystander reaction dataset as input to a deep-learning model, BADNet, to predict failure occurrence. We tested different data labeling methods and learned how they affect model performance, achieving precisions above 90%. We discuss strategies (manual labelling, failure-vs-control, and failure-time) used to model bystander reactions and predict failure, and how this approach can be used in real-world robotic deployments to detect errors and improve robot performance. As part of this work, we also contribute with the “Bystander Affect Detection” (BAD) dataset of bystander reactions, supporting the development of better prediction models.

Vision research showed remarkable success in understanding our world, propelled by datasets of images and videos. Sensor data from radar, LiDAR and cameras supports research in robotics and autonomous driving for at least a decade. However, while visual sensors may fail in some conditions, sound has recently shown potential to complement sensor data. Simulated room impulse responses (RIR) in 3D apartment-models became a benchmark dataset for the community, fostering a range of audiovisual research. In simulation, depth is predictable from sound, by learning bat-like perception with a neural network. Concurrently, the same was achieved in reality by using RGB-D images and echoes of chirping sounds. Biomimicking bat perception is an exciting new direction but needs dedicated datasets to explore the potential. Therefore, we collected the BatVision dataset to provide large-scale echoes in complex real-world scenes to the community. We equipped a robot with a speaker to emit chirps and a binaural microphone to record their echoes. Synchronized RGB-D images from the same perspective provide visual labels of traversed spaces. We sampled modern US office spaces to historic French university grounds, indoor and outdoor with large architectural variety. This dataset will allow research on robot echolocation, general audio-visual tasks and sound phænomena unavailable in simulated data. We show promising results for audio-only depth prediction and show how state-of-the-art work developed for simulated data can also succeed on our dataset. Project page: https://amandinebtto.github.io/Batvision-Dataset/

Affordable 3D scanners often produce sparse and non-uniform point clouds that negatively impact downstream applications in robotic systems. While existing point cloud upsampling architectures have demonstrated promising results on standard benchmarks, they tend to experience significant performance drops when the test data have different distributions from the training data. To address this issue, this paper proposes a test-time adaption approach to enhance model generality of point cloud upsampling. The proposed approach leverages meta-learning to explicitly learn network parameters for test-time adaption. Our method does not require any prior information about the test data. During meta-training, the model parameters are learned from a collection of instance-level tasks, each of which consists of a sparse-dense pair of point clouds from the training data. During meta-testing, the trained model is fine-tuned with a few gradient updates to produce a unique set of network parameters for each test instance. The updated model is then used for the final prediction. Our framework is generic and can be applied in a plug-and-play manner with existing backbone networks in point cloud upsampling. Extensive experiments demonstrate that our approach improves the performance of state-of-the-art models.

Deployable structures provide adaptability and versatility for applications such as temporary architectures, space structures, and biomedical devices. Jamming is a mechanical phenomenon with which dramatic changes in stiffness can be achieved by increasing the frictional and kinematic coupling between constituents in a structure by applying an external pressure. This study applies jamming, which has been primarily used in medium-scale soft robotics applications to large-scale deployable structures with components that are soft and compact during transport, but rigid upon deployment. It proposes a new jamming structure with a novel built-in actuation mechanism which enables high-performance at large scales: a composite beam made of rectangular segments along a cable which can be pre-tensioned and thus jammed. Two theoretical models are developed to provide insights into the mechanical behavior of the composite beams and predict their performance under loading. A scale model of a deployable bridge is built using the tension-based composite beams, and the bridge is deployed and assembled by air with a drone demonstrating the versatility and viability of the proposed approach for robotics applications.

We present TemporalStereo, a coarse-to-fine stereo matching network that is highly efficient, and able to effectively exploit the past geometry and context information to boost matching accuracy. Our network leverages sparse cost volume and proves to be effective when a single stereo pair is given. However, its peculiar ability to use spatio-temporal information across stereo sequences allows TemporalStereo to alleviate problems such as occlusions and reflective regions while enjoying high efficiency also in this latter case. Notably, our model - trained once with stereo videos - can run in both single-pair and temporal modes seamlessly. Experiments show that our network relying on camera motion is robust even to dynamic objects when running on videos. We validate TemporalStereo through extensive experiments on synthetic (SceneFlow, TartanAir) and real (KITTI 2012, KITTI 2015) datasets. Our model achieves state-of-the-art performance on any of these datasets.

We propose a framework that uses temporal logic specifications to predict and monitor the intent of a robotic agent through passive observations of its actions over time. Our approach uses a set of possible hypothesized intents specified as Büchi automata, obtained from translating temporal logic formulae. Based on observing the actions of the robot, we update the probabilities of each hypothesis using Bayes rule. Observations of robot actions provide strong evidence for its “immediate” short-term goals, whereas temporal logic specifications describe behaviors over a “never-ending” infinite time horizon. To bridge this gap, we use a two-level hierarchical monitoring approach. At the lower level, we track the immediate short-term goals of the robot which are modeled as atomic propositions in the temporal logic formalism. We apply our approach to predicting intent of human workers and thus their movements in an indoor space based on the publicly available THOR dataset. We show how our approach correctly labels each agent with their appropriate intents after relatively few observations while predicting their future actions accurately over longer time horizons.

This work presents a hierarchical framework for bipedal locomotion that combines a Reinforcement Learning (RL)-based high-level (HL) planner policy for the online generation of task space commands with a model-based low-level (LL) controller to track the desired task space trajectories. Different from traditional end-to-end learning approaches, our HL policy takes insights from the angular momentum-based linear inverted pendulum (ALIP) to carefully design the observation and action spaces of the Markov Decision Process (MDP). This simple yet effective design creates an insightful mapping between a low-dimensional state that effectively captures the complex dynamics of bipedal locomotion and a set of task space outputs that shape the walking gait of the robot. The HL policy is agnostic to the task space LL controller, which increases the flexibility of the design and generalization of the framework to other bipedal robots. This hierarchical design results in a learning-based framework with improved performance, data efficiency, and robustness compared with the ALIP model-based approach and state-of-the-art learning-based frameworks for bipedal locomotion. The proposed hierarchical controller is tested in three different robots, Rabbit, a five-link underactuated planar biped; Walker2D, a seven-link fully-actuated planar biped; and Digit, a 3D humanoid robot with 20 actuated joints. The trained policy naturally learns human-like locomotion behaviors and is able to effectively track a wide range of walking speeds while preserving the robustness and stability of the walking gait even under adversarial conditions.

This paper reports on Team Northeastern's Avatar system for telepresence, and our holistic approach to meet the ANA Avatar XPRIZE Final testing task requirements. The system features a dual-arm configuration with hydraulically actuated glove-gripper pair for haptic force feedback. Our proposed Avatar system was evaluated in the ANA Avatar XPRIZE Finals and completed all 10 tasks, scored 14.5 points out of 15.0, and received the 3rd Place Award. We provide the details of improvements over our first generation Avatar, covering manipulation, perception, locomotion, power, network, and controller design. We also extensively discuss the major lessons learned during our participation in the competition.

This paper studies a team coordination problem in a graph environment. Specifically, we incorporate “support” action which an agent can take to reduce the cost for its teammate to traverse some high cost edges. Due to this added feature, the graph traversal is no longer a standard multi-agent path planning problem. To solve this new problem, we propose a novel formulation that poses it as a planning problem in a joint state space: the joint state graph (JSG). Since the edges of JSG implicitly incorporate the support actions taken by the agents, we are able to now optimize the joint actions by solving a standard single-agent path planning problem in JSG. One main drawback of this approach is the curse of dimensionality in both the number of agents and the size of the graph. To improve scalability in graph size, we further propose a hierarchical decomposition method to perform path planning in two levels. We provide both theoretical and empirical complexity analyses to demonstrate the efficiency of our two algorithms.

Optimizing the morphologies and the controllers that adapt to various tasks is a critical issue in the field of robot design, aka. embodied intelligence. Previous works typically model it as a joint optimization problem and use search-based methods to find the optimal solution in the morphology space. However, they ignore the implicit knowledge of task-to-morphology mapping which can directly inspire robot design. For example, flipping heavier boxes tends to require more muscular robot arms. This paper proposes a novel and general differentiable task-inspired framework for contact-aware robot design called Task2Morph. We abstract task features highly related to task performance and use them to build a task-to-morphology mapping. Further, we embed the mapping into a differentiable robot design process, where the gradient information is leveraged for both the mapping learning and the whole optimization. The experiments are conducted on three scenarios, and the results validate that Task2Morph outperforms DiffHand, which lacks a task-inspired morphology module, in terms of efficiency and effectiveness.

In this paper, we study the problem of task-oriented grasp synthesis from partial point cloud data using an eye-in-hand camera configuration. In task-oriented grasp synthesis, a grasp has to be selected so that the object is not lost during manipulation, and it is also ensured that adequate force/moment can be applied to perform the task. We formalize the notion of a gross manipulation task as a constant screw motion (or a sequence of constant screw motions) to be applied to the object after grasping. Using this notion of task, and a corresponding grasp quality metric developed in our prior work, we use a neural network to approximate a function for predicting the grasp quality metric on a cuboid shape. We show that by using a bounding box obtained from the partial point cloud of an object, and the grasp quality metric mentioned above, we can generate a good grasping region on the bounding box that can be used to compute an antipodal grasp on the actual object. Our algorithm does not use any manually labeled data or grasping simulator, thus making it very efficient to implement and integrate with screw linear interpolation-based motion planners. We present simulation as well as experimental results that show the effectiveness of our approach. Website: https://irsl-sbu.github.io/Task-Oriented-Grasping-from-Point-Cloud-Representation/.

To perform household tasks, assistive robots receive commands in the form of user language instructions for tool manipulation. The initial stage involves selecting the intended tool (i.e., object grounding) and grasping it in a task-oriented manner (i.e., task grounding). Nevertheless, prior researches on visual-language grasping (VLG) focus on object grounding, while disregarding the fine-grained impact of tasks on object grasping. Task-incompatible grasping of a tool will inevitably limit the success of subsequent manipulation steps. Motivated by this problem, this paper proposes GraspCLIP, which addresses the challenge of task grounding in addition to object grounding to enable task-oriented grasp prediction with visual-language inputs. Evaluation on a custom dataset demonstrates that GraspCLIP achieves superior performance over established baselines with object grounding only. The effectiveness of the proposed method is further validated on an assistive robotic arm for grasping previously unseen kitchen tools given the task specification. Our presentation video is available at: https://www.youtube.com/watch?v=e1wfYQPeAXU.

Multi-object rearrangement is a crucial skill for service robots, and commonsense reasoning is frequently needed in this process. However, achieving commonsense arrangements requires knowledge about objects, which is hard to transfer to robots. Large language models (LLMs) are one potential source of this knowledge, but they do not naively capture information about plausible physical arrangements of the world. We propose LLM-GROP, which uses prompting to extract commonsense knowledge about semantically valid object configurations from an LLM and instantiates them with a task and motion planner in order to generalize to varying scene geometry. LLM-GROP allows us to go from natural-language commands to human-aligned object rearrangement in varied environments. Based on human evaluations, our approach achieves the highest rating while outperforming competitive baselines in terms of success rate while maintaining comparable cumulative action costs. Finally, we demonstrate a practical implementation of LLM-GROP on a mobile manipulator in real-world scenarios. Supplementary materials are available at: https://sites.google.com/view/llm-grop

Continuum robots suffer large deflections due to internal and external forces. Accurate modeling of their passive compliance is necessary for accurate environmental interaction, especially in scenarios where direct force sensing is not practical. This paper focuses on deriving analytic formulations for the compliance of continuum robots that can be modeled as Kirchhoff rods. Compared to prior works, the approach presented herein is not subject to the constant-curvature assumptions to derive the configuration space compliance, and we do not rely on computationally-expensive finite difference approximations to obtain the task space compliance. Using modal approximations over curvature space and Lie group integration, we obtain closed-form expressions for the task and configuration space compliance matrices of continuum robots, thereby bridging the gap between constant-curvature analytic formulations of configuration space compliance and variable curvature task space compliance. We first present an analytic expression for the compliance of a single Kirchhoff rod. We then extend this formulation for computing both the task space and configuration space compliance of a tendon-actuated continuum robot. We then use our formulation to study the tradeoffs between computation cost and modeling accuracy as well as the loss in accuracy from neglecting the Jacobian derivative term in the compliance model. Finally, we experimentally validate the model on a tendon-actuated continuum segment, demonstrating the model's ability to predict passive deflections with error below 11.5% percent of total arc length.

This paper proposes a task planning and motion control framework that generates task plans for a linear temporal logic specification (LTL), which are then executed using a task-space constrained motion controller and a local task planner that overcomes local minima. We propose a new encoding for task specifications, directly in the task-space, as constraints of a mixed-integer linear program that can be used with off-the-shelf LTL linear encoding. We apply our framework to plan and execute trajectories for a free-flying robot and show that the task plan is accomplished without collisions, even in the presence of unexpected moving obstacles that are not considered in the planning phase, while control signal constraints are satisfied. To evaluate the local minima avoidance, we compare the local task planner with a sampling-based motion planner, and the results show a smoother trajectory with a faster execution and less total planning time when using our framework. Last, our framework scaled well with a longer LTL specification, as opposed to automata-based frameworks that usually suffer with the curse of the dimensionality.

We address the Warehouse Servicing Problem (WSP) in automated warehouses, which use teams of mobile robots to move products from shelves to packaging stations. Given a list of products, the WSP amounts to finding a motion plan which brings every product on the list from a shelf to a packaging station within a given time limit. The WSP consists of four subproblems, namely, deciding where to source and deposit a product (task formulation), who should transport each product (task assignment) and when (scheduling) and how (motion planning). These problems are NP-Hard individually and made more challenging by their interdependence. The difficulty of the WSP is compounded by the scale of automated warehouses, which use teams of hundreds of agents to transport thousands of products. In this paper, we present Contract-based Cyclic Motion Planning (CCMP), a novel contract-based methodology for solving the WSP at scale. CCMP decomposes a warehouse into a set of traffic system components. By assigning each component a contract which describes the traffic flows it can support, CCMP can generate a traffic flow which satisfies a given WSP instance. CCMP then uses a novel motion planner to transform this traffic flow into a motion plan for a team of robots. Evaluation shows that CCMP can solve WSP instances taken from real industrial scenarios with up to 1 million products while outperforming other methodologies for solving the WSP by up to 2.9×.

In this paper, a comprehensive solution for enabling unmanned aerial vehicle (UAV) to autonomously fly through complex and dynamic environments is proposed. Moving objects all have unique property information, we propose a method that utilizes deep learning for 3D dynamic environment perception, while taking into account limitations in computing resources. For safer dynamic avoidance, we first model the dynamic target and integrate it into a static grid occupancy map, and then construct a gradient field based on its attribute information. To achieve autonomous UAV flight in dynamic environments, we design an adaptive planning method based on gradient optimization, which achieves significant computational savings by autonomously adjusting the planning frequency and using manually constructed gradients instead of maintaining a signed distance field (SDF). We have integrated the above approach into a customised quadrotor system and thoroughly tested it in real-world, verifying its flexibility to handle multiple objects with variable speed motion in complex enviroment.

In this paper, we propose a simple method, termed TWO, of directly closing the loop for LiDAR odometry. TWO suggests assigning high weights to the LIDAR observations corresponding to the old parts of the map; since these parts are built with the low-drift poses from the early odometry and can help drag the drifted odometry back to the correct global position when the LiDAR scans the points of these parts again. Also, we present the method of checking the consistency of the plane normal to address the two-side problem that may cause damage when using TWO. Moreover, we show that the proposed method is lightweight and needs little extra computation and storage space compared to the original odometry. The proposed TWO is integrated into the state-of-the-art LiDAR odometry A-LOAM and LiDAR-inertial odometry FAST-LIO2, and it is tested thoroughly on five public datasets and our private handheld dataset. The experiments show that the TWO can effectively help these two methods directly close most loops and produce localization results with apparently lower drifts.