ICRA 2024 论文列表

As a promising fashion for visual localization, scene coordinate regression (SCR) has seen tremendous progress in the past decade. Most recent methods usually adopt neural networks to learn the mapping from image pixels to 3D scene coordinates, which requires a vast amount of annotated training data. We propose to leverage Neural Radiance Fields (NeRF) to generate training samples for SCR. Despite NeRF’s efficiency in rendering, many of the rendered data are polluted by artifacts or only contain minimal information gain, which can hinder the regression accuracy or bring unnecessary computational costs with redundant data. These challenges are addressed in three folds in this paper: (1) A NeRF is designed to separately predict uncertainties for the rendered color and depth images, which reveal data reliability at the pixel level. (2) SCR is formulated as deep evidential learning with epistemic uncertainty, which is used to evaluate information gain and scene coordinate quality. (3) Based on the three arts of uncertainties, a novel view selection policy is formed that significantly improves data efficiency. Experiments on public datasets demonstrate that our method could select the samples that bring the most information gain and promote the performance with the highest efficiency.

With the rise in consumer depth cameras, a wealth of unlabeled RGB-D data has become available. This prompts the question of how to utilize this data for geometric reasoning of scenes. While many RGB-D registration methods rely on geometric and feature-based similarity, we take a different approach. We use cycle-consistent keypoints as salient points to enforce spatial coherence constraints during matching, improving correspondence accuracy. Additionally, we introduce a novel pose block that combines a GRU recurrent unit with transformation synchronization, blending historical and multi-view data. Our approach surpasses previous self-supervised registration methods on ScanNet and 3DMatch, even outperforming some older supervised methods. We also integrate our components into existing methods, showing their effectiveness.

Non-flat surfaces pose difficulties for robots operating in unstructured environments. Reconstructions of uneven surfaces may only be partially possible due to non-compliant end-effectors and limitations on vision systems such as transparency, reflections, and occlusions. This study achieves blind surface reconstruction by harnessing the robotic manipulator’s kinematic data and a compliant tactile sensing module, which incorporates inertial, magnetic, and pressure sensors. The module’s flexibility enables us to estimate contact positions and surface normals by analyzing its deformation during interactions with unknown objects. While previous works collect only positional information, we include the local normals in a geometrical approach to estimate curvatures between adjacent contact points. These parameters then guide a spline-based patch generation, which allows us to recreate larger surfaces without an increase in complexity while reducing the time-consuming step of probing the surface. Experimental validation demonstrates that this approach outperforms an off-the-shelf vision system in estimation accuracy. Moreover, this compliant haptic method works effectively even when the manipulator’s approach angle is not aligned with the surface normals, which is ideal for unknown non-flat surfaces.

Radar offers the advantage of providing additional physical properties related to observed objects. In this study, we design a physical-enhanced radar-inertial odometry system that capitalizes on the Doppler velocities and radar cross-section information. The filter for static radar points, correspondence estimation, and residual functions are all strengthened by integrating the physical properties. We conduct experiments on both public datasets and our self-collected data, with different mobile platforms and sensor types. Our quantitative results demonstrate that the proposed radar-inertial odometry system outperforms alternative methods using the physical-enhanced components. Our findings also reveal that using the physical properties results in fewer radar points for odometry estimation, but the performance is still guaranteed and even improved, thus aligning with the "less is more" principle.

Cataract is the leading cause of blindness worldwide with an increasing number of patients due to changing demographics, making automation an important part in future surgical treatment. In this work, we focus on a substep of cataract surgery, the Continuous Curvilinear Capsulorhexis (CCC). With a high complexity, this task is an ideal candidate for Reinforcement Learning (RL) in simulation. First, we present an interactive and physically realistic simulation based on the Finite Element Method (FEM) that mimics the tearing behavior of soft tissue during CCC. Then, we train and evaluate RL models in simulation, demonstrating that the trained policies can complete the CCC in 85% of cases. We also show that applying domain randomization techniques make the policy more robust against changes in geometrical and biomechanical boundary conditions.

Humans have a remarkable ability to fluently engage in joint collision avoidance in crowded navigation tasks despite the complexities and uncertainties inherent in human behavior. Underlying these interactions is a mutual understanding that (i) individuals are prosocial, that is, there is equitable responsibility in avoiding collisions, and (ii) individuals should behave legibly, that is, move in a way that clearly conveys their intent to reduce ambiguity in how they intend to avoid others. Toward building robots that can safely and seamlessly interact with humans, we propose a general robot trajectory planning framework for synthesizing legible and proactive behaviors and demonstrate that our robot planner naturally leads to prosocial interactions. Specifically, we introduce the notion of a markup factor to incentivize legible and proactive behaviors and an inconvenience budget constraint to ensure equitable collision avoidance responsibility. We evaluate our approach against well-established multi-agent planning algorithms and show that using our approach produces safe, fluent, and prosocial interactions. We demonstrate the real-time feasibility of our approach with human-in-the-loop simulations. Project page can be found at https://uw-ctrl.github.io/phri/.

Advancements in Automated Driving Systems (ADSs) have enabled the achievement of a certain level of autonomy while commuting in a car. However, emergency and high-speed maneuvers still arise as significant challenges for ADSs due to the intrinsic nonlinearity and fast-paced behavior of such events. These maneuvers are a distinctive feature within the recently established motorsport discipline of Autonomous Racing (AR). In this work, we explore the use of Learning-based Model Predictive Control (LMPC) to address possible model mismatches of the first principles model in high-speed racing. To this end, a Model Predictive Contouring Control (MPCC) (a specific formulation of the standard Model Predictive Control, MPC) is formulated, and a Neural Network (NN) that leverages the use of Feedforward and Recurrent layers is employed to learn the errors of the first principles model. By combining the NN with the first principles model, the LMPC is born, capable of accurately predicting the future with a computational effort compatible with real-time feasibility, effectively handling the vehicle at its limits. Furthermore, the controller can adapt to changing environments by training the NN during the race. The MPCC (formulation without the NN) is deployed on a real autonomous formula student racing car showing an improvement of 16 % in mean lap times across the same track between a common geometric controller. The LMPC is analyzed in a high-fidelity simulator, achieving an improvement of 8.9 % in mean lap times when compared to the MPCC.

Perception modules are integral in many modern autonomous systems, but their accuracy can be subject to the vagaries of the environment. In this paper, we propose a learning-based approach that can automatically characterize the error of a perception module from data and use this for safe control. The proposed approach constructs an inverse perception contract (IPC) which generates a set that contains the ground-truth value that is being estimated by the perception module, with high probability. We apply the proposed approach to study a vision pipeline deployed on a quadcopter. With the proposed approach, we successfully constructed an IPC for the vision pipeline. We then designed a control algorithm that utilizes the learned IPC, with the goal of landing the quadcopter safely on a landing pad. Experiments show that with the learned IPC, the control algorithm safely landed the quadcopter despite the error from the perception module, while the baseline algorithm without using the learned IPC failed to do so.

The trade-off between computation time and path optimality is a key consideration in motion planning algorithms. While classical sampling based algorithms fall short of computational efficiency in high dimensional planning, learning based methods have shown great potential in achieving time efficient and optimal motion planning. The SOTA learning based motion planning algorithms utilize paths generated by sampling based methods as expert supervision data and train networks via regression techniques. However, these methods often overlook the important multimodal property of the optimal paths in the training set, making them incapable of finding good paths in some scenarios. In this paper, we propose a Multimodal Neuron Planner (MNP) based on the mixture density networks that explicitly takes into account the multimodality of the training data and simultaneously achieves time efficiency and path optimality. For environments represented by point clouds, MNP first efficiently compresses point clouds into a latent vector by encoding networks that are suitable for processing point clouds. We then design multimodal planning networks which enables MNP to learn and predict multiple optimal solutions. Simulation results show that our method outperforms SOTA learning based method MPNet and advanced sampling based methods IRRT* and BIT*.

Model Predictive Control lacks the ability to escape local minima in nonconvex problems. Furthermore, in fast-changing, uncertain environments, the conventional warmstart, using the optimal trajectory from the last timestep, often falls short of providing an adequately close initial guess for the current optimal trajectory. This can potentially result in convergence failures and safety issues. Therefore, this paper proposes a framework for learning-aided warmstarts of Model Predictive Control algorithms. Our method leverages a neural network based multimodal predictor to generate multiple trajectory proposals for the autonomous vehicle, which are further refined by a sampling-based technique. This combined approach enables us to identify multiple distinct local minima and provide an improved initial guess. We validate our approach with Monte Carlo simulations of traffic scenarios.

Maneuverable tether-net systems launched from an unmanned spacecraft offer a promising solution for the active removal of large space debris. Guaranteeing the successful capture of such space debris is dependent on the ability to reliably maneuver the tether-net system – a flexible, many-DoF (thus complex) system – for a wide range of launch scenarios. Here, scenarios are defined by the relative location of the debris with respect to the chaser spacecraft. This paper represents and solves this problem as a hierarchically decentralized implementation of robotic trajectory planning and control and demonstrates the effectiveness of the approach when applied to two different tether-net systems, with 4 and 8 maneuverable units (MUs), respectively. Reinforcement learning (policy gradient) is used to design the centralized trajectory planner that, based on the relative location of the target debris at the launch of the net, computes the final aiming positions of each MU, from which their trajectory can be derived. Each MU then seeks to follow its assigned trajectory by using a decentralized PID controller that outputs the MU’s thrust vector and is informed by noisy sensor feedback (for realism) of its relative location. System performance is assessed in terms of capture success and overall fuel consumption by the MUs. Reward shaping and surrogate models are used to respectively guide and speed up the RL process. Simulation-based experiments show that this approach allows the successful capture of debris at fuel costs that are notably lower than nominal baselines, including in scenarios where the debris is significantly off-centered compared to the approaching chaser spacecraft.

Bio-inspired neural networks have the potential to advance our understanding of neural computation and improve the state-of-the-art of AI systems. Bio-electrical synapses directly transmit neural signals, by enabling fast current flow between neurons. In contrast, bio-chemical synapses transmit neural signals indirectly, through neurotransmitters. Prior work showed that interpretable dynamics for complex robotic control, can be achieved by using chemical synapses, within a sparse, bio-inspired architecture, called Neural Circuit Policies (NCPs). However, a comparison of these two synaptic models, within the same architecture, remains an unexplored area. In this work we aim to determine the impact of using chemical synapses compared to electrical synapses, in both sparse and all-to-all connected networks. We conduct experiments with autonomous lane-keeping through a photorealistic autonomous driving simulator to evaluate their performance under diverse conditions and in the presence of noise. The experiments highlight the substantial influence of the architectural and synaptic-model choices, respectively. Our results show that employing chemical synapses yields noticeable improvements compared to electrical synapses, and that NCPs lead to better results in both synaptic models.

Legged robots have the potential to traverse complex terrain and access confined spaces beyond the reach of traditional platforms thanks to their ability to carefully select footholds and flexibly adapt their body posture while walking. However, robust deployment in real-world applications is still an open challenge. In this paper, we present a method for legged locomotion control using reinforcement learning and 3D volumetric representations to enable robust and versatile locomotion in confined and unstructured environments. By employing a two-layer hierarchical policy structure, we exploit the capabilities of a highly robust low-level policy to follow 6D commands and a high-level policy to enable three-dimensional spatial awareness for navigating under overhanging obstacles. Our study includes the development of a procedural terrain generator to create diverse training environments. We present a series of experimental evaluations in both simulation and real-world settings, demonstrating the effectiveness of our approach in controlling a quadruped robot in confined, rough terrain. By achieving this, our work extends the applicability of legged robots to a broader range of scenarios.

To gently grasp objects, robots need to balance generating enough friction yet avoiding too much force that could damage the object. In practice, the force regulation is challenging to implement since it requires knowledge of the friction coefficient, which can vary from object to object and even from grasp to grasp. Tactile sensing offers a window in the contact mechanics and provides information about friction. Notably touch can detect the precursor of the object slipping away from the grasp. To find this information, tactile sensors measure the deformation field of an artificial skin in both the normal and tangential direction. However, current approaches only react to slip and therefore react too late to perturbations. The object slips, inducing a failure of the grasp and damage. In this study, we introduce a method that uses machine-learning to anticipate slip by computing the so-called safety margin of the grasp. This safety margin represents the extra lateral force that maintains the contact away from the frictional limit. To find this value, we use a high-density camera-based tactile sensor to measure the 3D deformation of the surface via the movement of 82 colored markers. We trained a Convolutional Neural Network (CNN) to estimate the safety margin from the tactile images. Because it gives a distance to slip, the safety margin is a powerful metric for regulating grasp forces. As a testament of this effectiveness, we show that a simple proportional controller can robustly grasp a wide variety of objects. The results show that this control method outperforms slip detection methods, by reducing regrasp reaction times while decreasing grasping forces to 1-3 N.

This work stages Foosball as a versatile platform for advancing scientific research, particularly in the realm of robot learning. We present an automated Foosball table along with its corresponding simulated counterpart, showcasing a diverse range of challenges through example tasks within the Foosball environment. Initial findings are shared using a simple baseline approach. Foosball constitutes a versatile learning environment with the potential to yield cutting-edge research in various fields of artificial intelligence and machine learning, notably robust learning, while also extending its applicability to industrial robotics and automation setups. To transform our physical Foosball table into a research-friendly system, we augmented it with a 2 degrees of freedom kinematic chain to control the goalkeeper rod as an initial setup with the intention to be extended to the full game as soon as possible. Our experiments reveal that a realistic simulation is essential for mastering complex robotic tasks, yet translating these accomplishments to the real system remains challenging, often accompanied by a performance decline. This emphasizes the critical importance of research in this direction. In this concern, we spotlight the automated Foosball table as an invaluable tool, possessing numerous desirable attributes, to serve as a demanding learning environment for advancing robotics and automation research.

Modern warehouses process millions of unique objects which are often stored in densely packed containers. To automate tasks in this environment, a robot must be able to pick diverse objects from highly cluttered scenes. Real-world learning is a promising approach, but executing picks in the real world is time-consuming, can induce costly failures, and often requires extensive human intervention, which causes operational burden and limits the scope of data collection and deployments. In this work, we leverage interactive probes to visually evaluate grasps in clutter without fully executing picks, a capability we refer to as Interactive Visual Failure Prediction (IVFP). This enables autonomous verification of grasps during execution to avoid costly downstream failures as well as autonomous reward assignment, providing supervision to continuously shape and improve grasping behavior as the robot gathers experience in the real world, without constantly requiring human intervention. Through experiments on a Stretch RE1 robot, we study the effect that IVFP has on performance - both in terms of effective data throughput and success rate, and show that this approach leads to grasping policies that outperform policies trained with human supervision alone, while requiring significantly less human intervention. Code, datasets, and videos available at https://robo-ivfp.github.io

In this paper, we address the challenge of exploring unknown indoor environments using autonomous aerial robots with Size Weight and Power (SWaP) constraints. The SWaP constraints induce limits on mission time requiring efficiency in exploration. We present a novel exploration framework that uses Deep Learning (DL) to predict the most likely indoor map given the previous observations, and Deep Reinforcement Learning (DRL) for exploration, designed to run on modern SWaP constraints neural processors. The DL-based map predictor provides a prediction of the occupancy of the unseen environment while the DRL-based planner determines the best navigation goals that can be safely reached to provide the most information. The two modules are tightly coupled and run onboard allowing the vehicle to safely map an unknown environment. Extensive experimental and simulation results show that our approach surpasses state-of-the-art methods by 50-60% in efficiency, which we measure by the fraction of the explored space as a function of the trajectory length.

Advancements in robot learning for object manipulation have shown promising results, yet certain everyday objects remain challenging for robots to effectively interact with. This discrepancy arises from the fact that human-designed objects are optimized for human use rather than robot manipulation. To address this gap, we propose a framework to automatically design 3D printable adaptations that can be attached to hard-to-use objects, thus improving "robot ergonomics". Our learning-based framework formulates the adaptation design and control as a dual Markov decision process and is able to improve robot-object interactions for various robot end effectors and objects. We further validate our designs in the real world with a Franka Panda robot. Please see the supplementary video and https://object-adaptation.github.io for additional visualizations.

Tracking and catching moving objects is an important ability for robots in a dynamic world. Whilst some objects have highly predictable state evolution e.g., the ballistic trajectory of a tennis ball, reactive targets alter their behavior in response to motion of the manipulator. Reactive applications range from gently capturing living animals such as snakes or fish for biological investigations, to smoothly interacting with and assisting a person. Existing works for dynamic catching usually perform target prediction followed by planning, but seldom account for highly non-linear reactive behaviors. Alternatively, Reinforcement Learning (RL) based methods simply treat the target and its motion as part of the observation of the world-state, but perform poorly due to the weak reward signal. In this work, we blend the approach of an explicit, yet learned, target state predictor with RL. We further show how a tightly coupled predictor which ‘observes’ the state of the robot leads to significantly improved anticipatory action, especially with targets that seek to evade the robot following a simple policy. Experiments show that our method achieves an 86.4% (open plane area) and a 73.8% (room) success rate on evasive objects, outperforming monolithic reinforcement learning and other techniques. We also demonstrate the efficacy of our approach across varied targets and trajectories. All code, data, and additional videos are at this GitHub link: https://kl-research.github.io/dyncatch.