ICRA 2024 论文列表

A self-driving vehicle must understand its environment to determine the appropriate action. Traditional autonomy systems rely on object detection to find the agents in the scene. However, object detection assumes a discrete set of objects and loses information about uncertainty, so any errors compound when predicting the future behavior of those agents. Alternatively, dense occupancy grid maps have been utilized to understand free-space. However, predicting a grid for the entire scene is wasteful since only certain spatio-temporal regions are reachable and relevant to the self-driving vehicle. We present a unified, interpretable, and efficient autonomy framework that moves away from cascading modules that first perceive, then predict, and finally plan. Instead, we shift the paradigm to have the planner query occupancy at relevant spatio-temporal points, restricting the computation to those regions of interest. Exploiting this representation, we evaluate a candidate trajectory around key factors such as collision avoidance, comfort, and progress for safety and interpretability. Our approach achieves better highway driving quality than the state-of-the-art on high-fidelity closed-loop simulations.

Cooperative perception can effectively enhance individual perception performance by providing additional viewpoint and expanding the sensing field. Existing cooperation paradigms are either interpretable (result cooperation) or flexible (feature cooperation). In this paper, we propose the concept of query cooperation to enable interpretable instance-level flexible feature interaction. To specifically explain the concept, we propose a cooperative perception framework, termed QUEST, which let query stream flow among agents. The cross-agent queries are interacted via fusion for co-aware instances and complementation for individual unaware instances. Taking camera-based vehicle-infrastructure perception as a typical practical application scene, the experimental results on the real-world dataset, DAIR-V2X-Seq, demonstrate the effectiveness of QUEST and further reveal the advantage of the query cooperation paradigm on transmission flexibility and robustness to packet dropout. We hope our work can further facilitate the cross-agent representation interaction for better cooperative perception in practice.

Designing information-rich and space-efficient sensors is a key challenge for soft robotics, and crucial for the development of safe soft robots. Sensing and understanding the environmental interactions with a minimal footprint is especially important in the medical context, where portability and unhindered patient/user movement is a priority, to move towards personalized and decentralized healthcare solutions. In this work, a pulsating fluidic soft sensor (PFS) capable of determining location, pressure and contact area of press events is shown. The sensor relies on spatio-temporal resistance changes driven by a pulsating conductive fluid. The sensor demonstrates good repeatability and distinction of single and multiple press events, detecting single indents of sizes greater than 1 cm, forces larger than 2 N, and various locations across the sensor, as well as multiple indents spaced 2 cm apart. Furthermore, the sensor is demonstrated in two applications to detect foot placement and grip location. Overall, the sensor represents an improvement towards minimizing electronic hardware, and cost of the sensing solution, without sacrificing the richness of the sensing information in the field of soft fluidic sensors.

The prevailing grasp prediction methods predominantly rely on offline learning, overlooking the dynamic grasp learning that occurs during real-time adaptation to novel picking scenarios. These scenarios may involve previously unseen objects, variations in camera perspectives, and bin configurations, among other factors. In this paper, we introduce a novel approach, SSL-ConvSAC, that combines semi-supervised learning and reinforcement learning for online grasp learning. By treating pixels with reward feedback as labeled data and others as unlabeled, it efficiently exploits unlabeled data to enhance learning. In addition, we address the imbalance between labeled and unlabeled data by proposing a contextual curriculum-based method. We ablate the proposed approach on real-world evaluation data and demonstrate promise for improving online grasp learning on bin picking tasks using a physical 7-DoF Franka Emika robot arm with a suction gripper. Video: https://youtu.be/OAro5pg8I9U

This study presents the Prosthetic Upper-Limb Sensory Enhancement (PULSE), a novel dual feedback device completely integrated into a prosthetic socket. The core of the system includes two compact vibrotactile actuators and two silicone chambers in contact with the user’s skin. These components provide high-frequency tactile cues for initial contact and surface information (e.g. texture) as well as pressure stimuli related to grasping force. Ten able-bodied participants and one subject with limb loss validated the system, accomplishing an object discrimination task in two different modalities (with and without the feedback). Standardized questionnaires evaluate users’ satisfaction and workload, enabling a systematic and robust device assessment. The results show that the PULSE device enhanced performance compared to its absence without causing discomfort for a prosthetic user and able-bodied participants. The findings highlight the potential of dual haptic feedback to enhance sensory perception in prosthetic applications and offer valuable insights for future prosthetic design.

The utilization of Large Language Models (LLMs) within the realm of reinforcement learning, particularly as planners, has garnered a significant degree of attention in recent scholarly literature. However, a substantial proportion of existing research predominantly focuses on planning models for robotics that transmute the outputs derived from perception models into linguistic forms, thus adopting a ‘pure-language’ strategy. In this research, we propose a hybrid End-to-End learning framework for autonomous driving by combining basic driving imitation learning with LLMs based on multi-modality prompt tokens. Instead of simply converting perception results from the separated train model into pure language input, our novelty lies in two aspects. 1) The end-to-end integration of visual and LiDAR sensory input into learnable multi-modality tokens, thereby intrinsically alleviating description bias by separated pre-trained perception models. 2) Instead of directly letting LLMs drive, this paper explores a hybrid setting of letting LLMs help the driving model correct mistakes and complicated scenarios. The results of our experiments suggest that the proposed methodology can attain driving scores of 49.21%, coupled with an impressive route completion rate of 91.34% in the offline evaluation conducted via CARLA. These performance metrics are comparable to the most advanced driving models.

In recent years, reinforcement learning and imitation learning have shown great potential for controlling humanoid robots’ motion. However, these methods typically create simulation environments and rewards for specific tasks, resulting in the requirements of multiple policies and limited capabilities for tackling complex and unknown tasks. To overcome these issues, we present a novel approach that combines adversarial imitation learning with large language models (LLMs). This innovative method enables the agent to learn reusable skills with a single policy and solve zero-shot tasks under the guidance of LLMs. In particular, we utilize the LLM as a strategic planner for applying previously learned skills to novel tasks through the comprehension of task-specific prompts. This empowers the robot to perform the specified actions in a sequence. To improve our model, we incorporate codebook-based vector quantization, allowing the agent to generate suitable actions in response to unseen textual commands from LLMs. Furthermore, we design general reward functions that consider the distinct motion features of humanoid robots, ensuring the agent imitates the motion data while maintaining goal orientation without additional guiding direction approaches or policies. To the best of our knowledge, this is the first framework that controls humanoid robots using a single learning policy network and LLM as a planner. Extensive experiments demonstrate that our method exhibits efficient and adaptive ability in complicated motion tasks.

While reinforcement learning (RL) attracts increasing research attention, maximizing the return while keeping the agent safe at the same time remains an open problem. Motivated to address this challenge, this work proposes a new Fast and Safe Policy Optimization (FSPO) algorithm, which consists of three steps: the first step involves reward improvement update, the second step projects the policy to the neighborhood of the baseline policy to accelerate the optimization process, and the third step addresses the constraint violation by projecting the policy back onto the constraint set. Such a projection-based optimization can improve the convergence and learning performance. Unlike many existing works that require convex approximations for the objectives and constraints, this work exploits a first-order method to avoid expensive computations and high dimensional issues, enabling fast and safe policy optimization, especially for challenging tasks. Numerical simulation and physical experiments demonstrate that FSPO outperforms existing methods in terms of safety guarantees and task completion rate.

Multi-task reinforcement learning could enable robots to scale across a wide variety of manipulation tasks in homes and workplaces. However, generalizing from one task to another and mitigating negative task interference still remains a challenge. Addressing this challenge by successfully sharing information across tasks will depend on how well the structure underlying the tasks is captured. In this work, we introduce our new architecture, Projected Task-Specific Layers (PTSL), that leverages a common policy with dense task-specific corrections through task-specific layers to better express shared and variable task information. We then show that our model outperforms the state of the art on the MT10 and MT50 benchmarks of Meta-World consisting of 10 and 50 goal-conditioned tasks for a Sawyer arm.

Internal limiting membrane (ILM) peeling is a vital vitreoretinal surgery procedure. However, due to the thickness of just 1-2 micrometers and the intricacies associated with its varying density and adhesion, the difficulty of manipulation exceeds the physiological limits of human perception and operation. Surgical robot is characterized by high precision and stability. However, navigating intricate intraocular environments and handling minuscule high-precision areas remain enormous challenges. These include issues of uneven lighting, field-of-view loss, and motion blur. This paper proposed a perception method named ‘Multimodal Surgical Process Recognition based on Domain Knowledge and Segmentation (MSPR-DKS),’ designed to address these challenges and provide input for the precise control of robots. Moreover, a comprehensive dataset focused on ILM peeling during macular hole surgeries was established. Experimental results underscore the efficacy of this approach, with segmentation accuracies exceeding 99.27% for instruments and macular holes and an average accuracy of 98.97% in recognizing surgical processes. This study paves the way for leveraging domain knowledge and image segmentation to improve robot-assisted manipulation of soft tissues in ophthalmology.

In this work, we focus on object search tasks within unexplored environments. We introduce a framework centered around the Probable Object Location (POLo) score. Utilizing a 3D object probability map, the POLo score allows the agent to make data-driven decisions for efficient object search. We further enhance the framework’s practicality by introducing POLoNet, a neural network trained to approximate the computationally-intensive POLo score. Our approach addresses critical limitations of both end-to-end reinforcement learning methods, which suffer from memory decay over long-horizon tasks, and traditional map-based methods that neglect visibility constraints. Our experiments, involving the first phase of the Open-Vocabulary Mobile Manipulation (OVMM) 2023 challenge, demonstrate that an agent equipped with POLoNet significantly outperforms a range of baseline methods, including end-to-end RL techniques and prior map-based strategies. To provide a comprehensive evaluation, we introduce new performance metrics that offer insights into the efficiency and effectiveness of various agents in object goal navigation.

The sense of touch is essential for robots to perform various daily tasks. Artificial Neural Networks have shown significant promise in advancing robotic tactile learning. However, due to the changing of tactile data distribution as robots encounter new tasks, ANN-based robotic tactile learning suffers from catastrophic forgetting. To solve this problem, we introduce a novel continual learning (CL) framework called the Probabilistic Spiking Neural Network with Variational Continual Learning (PSNN-VCL). In this framework, PSNN introduces uncertainty during spike emission and can apply fast Variational Inference by optimizing the uncertainty through backpropagation, which significantly reduces the required model parameters for VCL. We establish a robotic tactile CL benchmark using publicly available datasets to evaluate our method. Experimental results demonstrated that, compared to other CL methods, PSNN-VCL not only achieves superior performance in terms of widely used CL metrics but also achieves at least a 50% reduction in model parameters on the robotic tactile CL benchmark.

Autonomous mobile robots require predictions of human motion to plan a safe trajectory that avoids them. Because human motion cannot be predicted exactly, future trajectories are typically inferred from real-world data via learning-based approximations. These approximations provide useful information on the pedestrian’s behavior, but may deviate from the data, which can lead to collisions during planning. In this work, we introduce a joint prediction and planning framework, Partitioned Scenario Replay (PSR), that stores and partitions previously observed human trajectories, referred to as scenarios. During planning, scenarios observed in similar situations are reintroduced (or replayed) as motion predictions. By sampling real data and by building on scenario optimization and predictive control, the planner provides probabilistic collision avoidance guarantees in the real-world. Relying on this guarantee to remain safe, PSR can incrementally improve its prediction and planning performance online. We demonstrate our approach on a mobile robot navigating around pedestrians.

When a mobile robot autonomously explores an indoor space to produce a localization and navigation map, it is important to create both a stable pose graph and a high-quality occupancy map that covers all the navigable areas. In this work, we propose a novel probabilistic active loop closure framework which attempts to maximally reduce pose graph uncertainty during exploration and improves occupancy map quality. We calculate a probabilistic reward of getting a loop closure at any pose on a pose graph, which considers both how much pose graph uncertainty would be reduced by getting a loop closure there, and the robot’s travel cost to navigate to that pose. By choosing poses that provide the largest rewards, we can maximally reduce pose graph uncertainty while avoiding long travel times. The effectiveness of the method is illustrated through on-device testing in various floor plans.

Current state-of-the-art autonomous driving vehicles mainly rely on each individual sensor system to perform perception tasks. Such a framework’s reliability could be limited by occlusion or sensor failure. To address this issue, more recent research proposes using vehicle-to-vehicle (V2V) communication to share perception information with others. However, most relevant works focus only on cooperative detection and leave cooperative tracking an underexplored research field. A few recent datasets, such as V2V4Real, provide 3D multi-object cooperative tracking benchmarks. However, their proposed methods mainly use cooperative detection results as input to a standard single-sensor Kalman Filter-based tracking algorithm. In their approach, the measurement uncertainty of different sensors from different connected autonomous vehicles (CAVs) may not be properly estimated to utilize the theoretical optimality property of Kalman Filter-based tracking algorithms. In this paper, we propose a novel 3D multi-object cooperative tracking algorithm for autonomous driving via a differentiable multi-sensor Kalman Filter. Our algorithm learns to estimate measurement uncertainty for each detection that can better utilize the theoretical property of Kalman Filter-based tracking methods. The experiment results show that our algorithm improves the tracking accuracy by 17% with only 0.037x communication costs compared with the state-of-the-art method in V2V4Real. Our code and videos are available at the URL and the URL.

Collaborative manipulation task often requires negotiation using explicit or implicit communication. An important example is determining where to move when the goal destination is not uniquely specified, and who should lead the motion. This work is motivated by the ability of humans to communicate the desired destination of motion through back-and-forth force exchanges. Inherent to these exchanges is also the ability to dynamically assign a role to each participant, either taking the initiative or deferring to the partner’s lead. In this paper, we propose a hierarchical robot control framework that emulates human behavior in communicating a motion destination to a human collaborator and in responding to their actions. At the top level, the controller consists of a set of finite-state machines corresponding to different levels of commitment of the robot to its desired goal configuration. The control architecture is loosely based on the human strategy observed in the human-human experiments, and the key component is a real-time intent recognizer that helps the robot respond to human actions. We describe the details of the control framework, feature engineering and training process of the intent recognition. The proposed controller was implemented on a UR10e robot (Universal Robots) and evaluated through human studies. The experiments show that the robot correctly recognizes and responds to human input, communicates its intent clearly, and resolves conflict. We report success rates and draw comparisons with human-human experiments to demonstrate the effectiveness of the approach.

With the rapid development of autonomous driving systems, 3D object detection based on Bird’s Eye View (BEV) in road scenes has witnessed great progress over the past few years. As a road scene exhibits a part-whole hierarchy between the within objects and the scene itself, simple parts (e.g., roads, lane lines, vehicles and pedestrians) can be assembled into progressively more complex shapes to form a BEV representation of the whole road scene. Therefore, a BEV often has multiple levels of freedom on motion, i.e., the rotation and the moving shift of the whole BEV, and the random movements of objects (e.g., pedestrians and vehicles) inside the BEV. However, most of the current single-sensor or multi-sensor fusion-based BEV object detection methods have not yet taken into account capturing such multi-level motion in a BEV. To address this problem, we propose a product group equivariant object detection network framework that is equivariant with respect to multiple levels of symmetry groups based on multi-sensor fusion. The proposed framework extracts local equivariant features of objects in point clouds, while global equivariant features are extracted in both point clouds and images. Furthermore, the network learns diverse rotation-equivariant features and mitigates a significant amount of detection errors caused by rotations of BEV and objects inside a BEV, thereby further enhancing the performance of object detection. The experiment results show that the network architecture significantly improves object detection on mAP and NDS, respectively. In addition, in order to demonstrate the effectiveness of the proposed local-multi-global equivariant components, we conduct sufficient ablation experiments. The results show that the individual components are indispensable for the object detection performance improvement of the overall network architecture.

The prominence of embodied Artificial Intelligence (AI), which empowers robots to navigate, perceive, and engage within virtual environments, has attracted significant attention, owing to the remarkable advances in computer vision and large language models. Privacy emerges as a pivotal concern within the realm of embodied AI, as the robot accesses substantial personal information. However, the issue of privacy leakage in embodied AI tasks, particularly concerning reinforcement learning algorithms, has not received adequate consideration in research. This paper aims to address this gap by proposing an attack on the training process of the value-based algorithm and the gradient-based algorithm, utilizing gradient inversion to reconstruct states, actions, and supervisory signals. The choice of using gradients for the attack is motivated by the fact that commonly employed federated learning techniques solely utilize gradients computed based on private user data to optimize models, without storing or transmitting the data to public servers. Nevertheless, these gradients contain sufficient information to potentially expose private data. To validate our approach, we conducted experiments on the AI2THOR simulator and evaluated our algorithm on active perception, a prevalent task in embodied AI. The experimental results demonstrate the effectiveness of our method in successfully reconstructing all information from the data in 120 room layouts. Check our website for videos.

We are interested in studying sports with robots and starting with the problem of intercepting a projectile moving toward a robot manipulator equipped with a shield. To successfully perform this task, the robot needs to (i) detect the incoming projectile, (ii) predict the projectile’s future motion, (iii) plan a minimum-time rapid trajectory that can evade obstacles and intercept the projectile, and (iv) execute the planned trajectory. These four steps must be performed under the manipulator’s dynamic limits and extreme time constraints (≤ 350ms in our setting) to successfully intercept the projectile. In addition, we want these trajectories to be smooth to reduce the robot’s joint torques and the impulse on the platform on which it is mounted. To this end, we propose a kinodynamic motion planning framework that preprocesses smooth trajectories offline to allow real-time collision-free executions online. We present an end-to-end pipeline along with our planning framework, including perception, prediction, and execution modules. We evaluate our framework experimentally in simulation and show that it has a higher blocking success rate than the baselines. Further, we deploy our pipeline on a robotic system comprising an industrial arm (ABB IRB-1600) and an onboard stereo camera (ZED 2i), which achieves a 78% success rate in projectile interceptions.