ICRA 2024 论文列表

Coordinating heterogeneous robots is essential for autonomous multi-robot teaming. To execute a set of dependent tasks as quickly as possible, and to complete tasks that cannot be addressed by individual robots, it is necessary to form subteams that can collaboratively finish the tasks. It is also advantageous for robots to wait for teammates and tasks to become available in order to form better subteams or reduce the overall completion time. To enable both abilities, we introduce a new graph learning approach that formulates heterogeneous collaborative scheduling as a bipartite matching problem that maximizes a reward matrix learned via imitation learning. We design a novel graph attention transformer network (GATN) that represents the problem of collaborative scheduling as a bipartite graph, and integrates both local and global graph information to estimate the reward matrix using graph attention networks and transformers. By relaxing the constraint of one-to-one correspondence in bipartite matching, our approach allows multiple robots to address the same task as a subteam. Our approach also enables voluntary waiting by introducing an idle task that the robots can select to wait. Experimental results have shown that our approach well addresses heterogeneous collaborative scheduling with dynamic subteam formation and voluntary waiting, and outperforms the previous and baseline methods.

Manipulation of deformable Linear objects (DLOs), including iron wire, rubber, silk, and nylon rope, is ubiquitous in daily life. These objects exhibit diverse physical properties, such as Young’s modulus and bending stiffness. Such diversity poses challenges for developing generalized manipulation policies. However, previous research limited their scope to single-material DLOs and engaged in time-consuming data collection for the state estimation. In this paper, we propose a two-stage manipulation approach consisting of a material property (e.g., flexibility) estimation and policy learning for DLO insertion with reinforcement learning. Firstly, we design a flexibility estimation scheme that characterizes the properties of different types of DLOs. The ground truth flexibility data is collected in simulation to train our flexibility estimation module. During the manipulation, the robot interacts with the DLOs to estimate flexibility by analyzing their visual configurations. Secondly, we train a policy conditioned on the estimated flexibility to perform challenging DLO insertion tasks. Our pipeline trained with diverse insertion scenarios achieves an 85.6% success rate in simulation and 66.67% in real robot experiments. Please refer to our project page: https://lmeee.github.io/DLOInsert/

We investigate the active manipulation of objects using model-free and long-horizon DRL (Deep Reinforcement Learning) to achieve target shapes. Our proposed approach uses visual observations consisting of segmented images, to mitigate the sim-to-real gap. We address a long-horizon manipulation task requiring a sequence of accurate actions to achieve the target shapes using a robot arm with an RGB-D camera in eye-in-hand configuration, and an elongated, volumetric, elastoplastic object. We find similar objects in food, marine, and manufacturing domains. The aim is to actively manipulate the object into an arbitrary target shape using image observations. We trained a DRL agent using PPO (Proximal Policy Optimization) by running 768 parallel actors in simulation, for a total of 1,2M environment interactions, and tested this on 200 unseen target deformations. In three attempts, 82% of the trials achieved a greater than 90% overlap with the 200 target shapes. By relying on segmentation images as a visual observation space, we successfully transferred the agent to the real world without supplementary training. Our approach does not need any real-world manipulation examples nor fine-tuning in the real world. The robustness of our approach was demonstrated in simulation, and experimentally validated in the real world for specific manipulation tasks, achieving a 94.2% mean zero-shot overlap success rate on previously unseen target shapes.

Autonomous Dynamic System (DS)-based algorithms hold a pivotal and foundational role in the field of Learning from Demonstration (LfD). Nevertheless, they confront the formidable challenge of striking a delicate balance between achieving precision in learning and ensuring the overall stability of the system. In response to this substantial challenge, this paper introduces a novel DS algorithm rooted in neural network technology. This algorithm not only possesses the capability to extract critical insights from demonstration data but also demonstrates the capacity to learn a candidate Lyapunov energy function that is consistent with the provided demonstrations. The model presented in this paper employs a simplistic neural network architecture that excels in fulfilling a dual objective: optimizing accuracy while simultaneously preserving global stability. To comprehensively evaluate the effectiveness of the proposed algorithm, rigorous assessments are conducted using the LASA dataset, further reinforced by empirical validation through a robotic experiment.

Autonomous underwater vehicles often perform surveys that capture multiple views of targets in order to provide more information for human operators or automatic target recognition algorithms. In this work, we address the problem of choosing the most informative views that minimize survey time while maximizing classifier accuracy. We introduce a novel active perception framework for multi-view adaptive surveying and reacquisition using side scan sonar imagery. Our framework addresses this challenge by using a graph formulation for the adaptive survey task. We then use Graph Neural Networks (GNNs) to both classify acquired sonar views and to choose the next best view based on the collected data. We evaluate our method using simulated surveys in a high-fidelity side scan sonar simulator. Our results demonstrate that our approach is able to surpass the state-of-the-art in classification accuracy and survey efficiency. This framework is a promising approach for more efficient autonomous missions involving side scan sonar, such as underwater exploration, marine archaeology, and environmental monitoring.

Robots operating alongside humans often encounter unfamiliar environments that make autonomous task completion challenging. Though improving models and increasing dataset size can enhance a robot’s performance in unseen environments, data collection and model refinement may be impractical in every environment. Approaches that utilize human demonstrations through manual operation can aid in refinement and generalization, but often require significant data collection efforts to generate enough demonstration data to achieve satisfactory task performance. Interactive approaches allow for humans to provide correction to robot action in real time, but intervention policies are often based on explicit factors related to state and task understanding that may be difficult to generalize. Addressing these challenges, we train a lightweight interaction policy that allows robots to decide when to proceed autonomously or request expert assistance at estimated times of uncertainty. An implicit estimate of uncertainty is learned via evaluating the feature extraction capabilities of the robot’s visual navigation policy. By incorporating part-time human interaction, robots recover quickly from their mistakes, significantly improving the odds of task completion. Incorporating part-time interaction yields an increase in success of 0.38 with only a 0.3 expert interaction rate within the Habitat simulation environment using a simulated human expert. We further show success transferring this approach to a new domain with a real human expert, improving success from less than 0.1 with an autonomous agent to 0.92 with a 0.23 human interaction rate. This approach provides a practical means for robots to interact and learn from humans in real-world settings.

Learning strategic robot behavior—like that required in pursuit-evasion interactions—under real-world constraints is extremely challenging. It requires exploiting the dynamics of the interaction, and planning through both physical state and latent intent uncertainty. In this paper, we transform this intractable problem into a supervised learning problem, where a fully-observable robot policy generates supervision for a partially-observable one. We find that the quality of the supervision signal for the partially-observable pursuer policy depends on two key factors: the balance of diversity and optimality of the evader’s behavior, and the strength of the modeling assumptions in the fully-observable policy. We deploy our policy on a physical quadruped robot with an RGB-D camera on pursuit-evasion interactions in the wild. Despite all the challenges, the sensing constraints bring about creativity: the robot is pushed to gather information when uncertain, predict intent from noisy measurements, and anticipate in order to intercept.

Reinforcement learning (RL) for bipedal locomotion has recently demonstrated robust gaits over moderate terrains using only proprioceptive sensing. However, such blind controllers will fail in environments where robots must anticipate and adapt to local terrain, which requires visual perception. In this paper, we propose a fully-learned system that allows bipedal robots to react to local terrain while maintaining commanded travel speed and direction. Our approach first trains a controller in simulation using a heightmap expressed in the robot’s local frame. Next, data is collected in simulation to train a heightmap predictor, whose input is the history of depth images and robot states. We demonstrate that with appropriate domain randomization, this approach allows for successful sim-to-real transfer with no explicit pose estimation and no fine-tuning using real-world data. To the best of our knowledge, this is the first example of sim-to-real learning for vision-based bipedal locomotion over challenging terrains.

Collision Avoidance/Mitigation System (CAMS) for autonomous vehicles is a crucial technology that ensures the safety and reliability of autonomous driving systems. Conventional collision avoidance approaches struggle in complex and various scenarios by avoiding collisions based on rules for specific collision scenarios. This has led to learning-based methods using neural networks for adaptive collision avoidance. However, the approaches directly outputting control inputs through neural networks have drawbacks in interpretability and stability. To address these limitations, we propose a trajectory planning method for CAMS that combines deep reinforcement learning (DRL) and quintic polynomial (QP) trajectory planning. The proposed method determines the terminal state and confidence of the trajectory using DRL and plans a QP trajectory based on them. By utilizing the terminal state and confidence of the trajectory rather than direct control inputs as the output of the neural network, it generates a more realistic and continuous path. Moreover, this approach considers collision avoidance and mitigation in an integrated manner through the reward function of RL. Our experimental results demonstrate that the proposed method not only improves interpretability and stability compared to existing learning-based methods but also upholds performance in complex and various collision scenarios.

In autonomous driving and robotics, there is a growing interest in utilizing short-term historical data to enhance multi-camera 3D object detection, leveraging the continuous and correlated nature of input video streams. Recent work has focused on spatially aligning BEV-based features over timesteps. However, this is often limited as its gain does not scale well with long-term past observations. To address this, we advocate for supervising a model to predict objects’ poses given past observations, thus explicitly guiding to learn objects’ temporal cues. To this end, we propose a model called DAP (Detection After Prediction), consisting of a two-branch network: (i) a branch responsible for forecasting the current objects’ poses given past observations and (ii) another branch that detects objects based on the current and past observations. The features predicting the current objects from branch (i) is fused into branch (ii) to transfer predictive knowledge. We conduct extensive experiments with the large-scale nuScenes datasets, and we observe that utilizing such predictive information significantly improves the overall detection performance. Our model can be used plug-and-play, showing consistent performance gain.

In Human-Robot Interaction (HRI) scenarios, human factors like trust can greatly impact task performance and interaction quality. Recent research has confirmed that perceived robot proficiency is a major antecedent of trust. By making robots aware of their capabilities, we can allow them to choose when to perform low-confidence actions, thus actively controlling the risk of trust reduction. In this paper, we propose Self-Confidence through Observed Novel Experiences (SCONE), a policy to learn self-confidence from experience using semantic action embeddings. Using an assistive cooking setting, we show that the semantic aspect allows SCONE to learn self-confidence faster than existing approaches, while also achieving promising performance in simple instructions following. Finally, we share results from a pilot study with 31 participants, showing that such a self-confidence-aware policy increases capability-based human trust.

Deployment in hazardous environments requires robots to understand the risks associated with their actions and movements to prevent accidents. Despite its importance, these risks are not explicitly modeled by currently deployed locomotion controllers for legged robots. In this work, we propose a risk sensitive locomotion training method employing distributional reinforcement learning to consider safety explicitly. Instead of relying on a value expectation, we estimate the complete value distribution to account for uncertainty in the robot’s interaction with the environment. The value distribution is consumed by a risk metric to extract risk sensitive value estimates. These are integrated into Proximal Policy Optimization (PPO) to derive our method, Distributional Proximal Policy Optimization (DPPO). The risk preference, ranging from risk-averse to risk-seeking, can be controlled by a single parameter, which enables to adjust the robot’s behavior dynamically. Importantly, our approach removes the need for additional reward function tuning to achieve risk sensitivity. We show emergent risk sensitive locomotion behavior in simulation and on the quadrupedal robot ANYmal. Videos of the experiments and code are available at https://sites.google.com/leggedrobotics.com/risk-aware-locomotion.

Grasping is one of the most fundamental skills for humans to interact with objects. However, it remains a challenging problem for anthropomorphic hands, due to the lack of object affordance understanding and high-dimensional grasp planning. In this work, we propose an anthropomorphic hand grasping framework to learn realistic and reasonable grasps in cluttered scenes, which tackles the problem in three items: 1) graspable point segmentation; 2) hand grasp generation and 3) grasp optimization. Specifically, our method generates high-quality hand grasps efficiently without complete object models by learning graspable points, associated grasp configurations from observed point cloud in a parallel manner and optimizing predicted grasps based on hand-object contacts. Simulation experiments show that our model generates physical plausible grasps for the anthropomorphic hand effectively with over 70% success rate. Real-world experiments demonstrate that the model trained in simulation performs satisfactorily in real-world scenarios for unseen objects.

Quadrupedal robots have played a crucial role in various environments, from structured environments to complex harsh terrains, thanks to their agile locomotion ability. However, these robots can easily lose their locomotion functionality if damaged by external accidents or internal malfunctions. In this paper, we propose a novel deep reinforcement learning framework to enable a quadrupedal robot to walk with impaired joints. The proposed framework consists of three components: 1) a random joint masking strategy for simulating impaired joint scenarios, 2) a joint state estimator to predict an implicit status of current joint condition based on past observation history, and 3) progressive curriculum learning to allow a single network to conduct both normal gait and various joint-impaired gaits. We verify that our framework enables the Unitree’s Go1 robot to walk under various impaired joint conditions in real-world indoor and outdoor environments.

While instance segmentation models excel at object detection in satellite imagery, their performance drops when applied to slant-angled aerial images due to occlusion and scale variation. This is mainly caused by a lack of training data for such diverse viewpoints and scales. To address this limitation, we propose the Sim2Real-based Multi-Scale Context Mask-RCNN (MSC-RCNN) network, specifically designed for slant-angled aerial imagery. Sim2Real-based transfer learning is adapted to compensate for the limited availability of real-world slant-angle training data. A synthetic dataset is generated using Unreal Engine, detailing the methodology of replicating the real-world scene, for producing diverse slant-angle drone datasets with various weather conditions and backgrounds. The model leverages two distinct feature pyramid backbones, with one incorporating dilated convolutions to address large-scale objects and the other optimized for regular convolutions. Their outputs are fused to effectively detect objects across various scales and angles. Through experiments, it was demonstrated that incorporating this synthetic data significantly reduces reliance on real data while maintaining high mean Average Precision (mAP) scores. Compared to the baseline Mask R-CNN, the proposed approach with Sim2Real adaptation and the MSC-RCNN architecture achieves a remarkable 7.6% performance improvement in instance segmentation accuracy with only a 6% increase in model size. Code can be found at: https://github.com/MSC-RCNN

Learning manipulation from demonstration is a key way for humans to teach complex tasks. However, this domain mainly focuses on kinetic teaching, and does not consider imitation of interaction forces which is essential for more contact rich tasks. We propose a framework that enables robotic imitation of contact from human demonstration using a wearable finger-tip sensor. By developing a multi-modal sensor (providing both force and contact location) and robotic collection of simple training data of different motion primitives (tapping, rotation and translation), an LSTM-based model can be used to replicate motion from tactile demonstration only. To evaluate this approach, we explore the performance on increasingly complex testing data generated by a robot, and also demonstrate the full pipeline from human demonstration via the sensor used as a wearable device. This approach of using tactile sensing as a means of inferring the required robot motion paves the way for imitation of more contact-rich tasks, and enables imitation of tasks where the demonstration and imitation is performed with different body-schema.

This work presents a novel Learning Model Predictive Control (LMPC) strategy for autonomous racing at the handling limit that can iteratively explore and learn unknown dynamics in high-speed operational domains. We start from existing LMPC formulations and modify the system dynamics learning method. In particular, our approach uses a nominal, global, nonlinear, physics-based model with a local, linear, data-driven learning of the error dynamics. We conducted experiments in simulation and on 1/10th scale hardware, and deployed the proposed LMPC on a full-scale autonomous race car used in the Indy Autonomous Challenge (IAC) with closed loop experiments at the Putnam Park Road Course in Indiana, USA. The results show that the proposed control policy exhibits improved robustness to parameter tuning and data scarcity. Incremental and safety-aware exploration toward the limit of handling and iterative learning of the vehicle dynamics in high-speed domains is observed both in simulations and experiments.

Multi-task learning of deformable object manipulation is a challenging problem in robot manipulation. Most previous works address this problem in a goal-conditioned way and adapt goal images to specify different tasks, which limits the multi-task learning performance and can not generalize to new tasks. Thus, we adapt language instruction to specify deformable object manipulation tasks and propose a learning framework. We first design a unified Transformer-based architecture to understand multi-modal data and output picking and placing action. Besides, we have applied the visible connectivity graph to tackle nonlinear dynamics and complex configuration of the deformable object. Both simulated and real experiments have demonstrated that the proposed method is effective and can generalize to unseen instructions and tasks. Compared with the state-of-the-art method, our method achieves higher success rates (87.2% on average) and has a 75.6% shorter inference time. We also demonstrate that our method performs well in real-world experiments. Supplementary videos can be found at https://sites.google.com/view/language-deformable.

Cross-pose estimation between rigid objects is a fundamental building block for robotic applications. In this paper, we propose a new cross-pose estimation method that predicts correspondences on a set level as opposed to a point level. This contrasts methods that predict cross-pose from per-point correspondences, which can encounter optimization problems for objects with symmetries, since each point may have multiple valid correspondences. Our method, SCAlign, consists of a Set Correspondence Network (SCN) which predicts these sets and their correspondences, and an alignment module to compute their relative cross-pose. Taking point clouds of two objects as input, SCN predicts a set label for each point such that such that points that share a set label form a cross object correspondence. The alignment module then computes the cross-pose as the SE(3) transformation that aligns these set correspondences. We compare SCAlign against other cross-pose estimation baselines on a synthetically generated dataset, SynWidth, which contains randomly generated width-mate objects with symmetric or near-symmetric intercepts. SCAlign significantly outperforms the baselines on this challenging dataset. Additionally, we show that set correspondences can be leveraged to distinguish positive and negative matches between pegs and holes. Robot experiments further validate the practical application of this approach.