IROS 2024 论文列表

In this paper, we introduce a framework for imitation learning in navigation that enables policy learning from one-shot images without a physical robot and facilitates the transfer of this policy from simulation to reality. Utilizing Neural Radiance Fields (NeRF), our approach generates a simulated environment and simultaneously models expert behavior. This removes the necessity for a physical robot during both the expert teaching phase and the agent’s learning process, allowing for the application of policies learned within the NeRF simulation to real-world robots. We validate our method by demonstrating the navigation with an actual robot using the policy learned by our approach. Moreover, we present a method for adapting to changes in the robot configuration, such as camera parameters and robot dimensions, by simulating adjustments in the robot configuration throughout the learning and assessing its generalizability.

Interacting with the actual environment to acquire data is often costly and time-consuming in robotic tasks. Model-based offline reinforcement learning (RL) provides a feasible solution. On the one hand, it eliminates the requirements of interaction with the actual environment. On the other hand, it learns the transition dynamics and reward function from the offline datasets and generates simulated rollouts to accelerate training. Previous model-based offline RL methods adopt probabilistic ensemble neural networks (NN) to model aleatoric uncertainty and epistemic uncertainty. However, this results in a great increase in training time and computing resource requirements. Furthermore, these methods are easily disturbed by the accumulative errors of the environment dynamics models when simulating long-term rollouts. To solve the above problems, we propose an uncertainty-aware sequence modeling architecture called Environment Transformer. It models the probability distribution of the environment dynamics and reward function to capture aleatoric uncertainty and treats epistemic uncertainty as a learnable noise parameter. Benefiting from the accurate modeling of the transition dynamics and reward function, Environment Transformer can be combined with arbitrary planning, dynamics programming, or policy optimization algorithms for offline RL. In this case, we perform Conservative Q-Learning (CQL) to learn a conservative Q-function. Through simulation experiments, we demonstrate that our method achieves or exceeds state-of-the-art performance in widely studied offline RL benchmarks. Moreover, we show that Environment Transformer's simulated rollout quality, sample efficiency, and long-term rollout simulation capability are superior to those of previous model-based offline RL methods.

Integrating Large Language Models (LLMs) into robotics significantly enhances autonomous task planning. However, ensuring that multi-step task plans (action sequence) generated by LLMs comply with pre-defined safety constraints during planning and execution remains a challenge, limiting their adaptability in complex environments. To address this issue, a mechanism that can monitor and adjust the plan generated by the LLM-driven task planner and guide the motion planner to avoid potential risks during action execution is required. Therefore, this paper proposes a cross-layer sequence supervision mechanism. Specifically, we employ linear temporal logic syntax to express safety constraints and convert them into a set of nondeterministic Büchi automatons to build a cross-layer safety supervisor. For the task planning layer, the safety supervisor provides a closed-loop correction mechanism that can identify violations in the task plan in real time and guide LLM-driven planners to correct this plan to ensure compliance. For the motion planning layer, the safety supervisor introduces virtual "obstacle" information into the task plan to form the task plan tuple. Based on this plan tuple, the motion planner can proactively prevent unsafe behaviors during action execution. Extensive experimentation demonstrates significant improvements in safety with this cross-layer supervision mechanism, highlighting its potential to enhance LLM-driven robotic technology. Experiment details can be found in https://youtu.be/BDdSSEP6HJw.

This paper proposes an optimization-based control framework for the torque-controlled robot, which can satisfy the joint position, velocity, and acceleration constraints under a bounded jerk. The optimization filter is incorporated as a module to modify the nominal controller output to ensure joint constraints. To formulate the optimization problem as a QP, the torque optimization problem is converted to the jerk optimization problem using the augmented state, and the constraints are reformulated to be affine in the jerk. Here, the viable constraints are derived using the time-optimal braking policy to guarantee the feasibility of the QP. The proposed method was validated in simulation and with a 6-DOF robot manipulator.

Multi-Agent Reinforcement Learning (MARL) based Multi-Agent Path Finding (MAPF) has recently gained attention due to its efficiency and scalability. Several MARL-MAPF methods choose to use communication to enrich the information one agent can perceive. However, existing works still struggle in structured environments with high obstacle density and a high number of agents. To further improve the performance of the communication-based MARL-MAPF solvers, we propose a new method, Ensembling Prioritized Hybrid Policies (EPH). We first propose a selective communication block to gather richer information for better agent coordination within multi-agent environments and train the model with a Q-learning-based algorithm. We further introduce three advanced inference strategies aimed at bolstering performance during the execution phase. First, we hybridize the neural policy with single-agent expert guidance for navigating conflict-free zones. Secondly, we propose Q value-based methods for prioritized resolution of conflicts as well as deadlock situations. Finally, we introduce a robust ensemble method that can efficiently collect the best out of multiple possible solutions. We empirically evaluate EPH in complex multi-agent environments and demonstrate competitive performance against state-of-the-art neural methods for MAPF. We open-source our code at https://github.com/ai4co/eph-mapf.

Event cameras are increasingly popular in robotics due to beneficial features such as low latency, energy efficiency, and high dynamic range. Nevertheless, their downstream task performance is greatly influenced by the optimization of bias parameters. These parameters, for instance, regulate the necessary change in light intensity to trigger an event, which in turn depends on factors such as the environment lighting and camera motion. This paper introduces feedback control algorithms that automatically tune the bias parameters through two interacting methods: 1) An immediate, on-the-fly fast adaptation of the refractory period, which sets the minimum interval between consecutive events, and 2) if the event rate exceeds the specified bounds even after changing the refractory period repeatedly, the controller adapts the pixel bandwidth and event thresholds, which stabilizes after a short period of noise events across all pixels (slow adaptation). Our evaluation focuses on the visual place recognition task, where incoming query images are compared to a given reference database. We conducted comprehensive evaluations of our algorithms’ adaptive feedback control in real-time. To do so, we collected the QCR-Fast-and-Slow dataset that contains DAVIS346 event camera streams from 366 repeated traversals of a Scout Mini robot navigating through a 100 meter long indoor lab setting (totaling over 35km distance traveled) in varying brightness conditions with ground truth location information. Our proposed feedback controllers result in superior performance when compared to the standard bias settings and prior feedback control methods. Our findings also detail the impact of bias adjustments on task performance and feature ablation studies on the fast and slow adaptation mechanisms.

In soft tissue surgeries, such as tumor resections, achieving precision is of utmost importance. Surgeons conventionally achieve this precision through intraoperative adjustments to the cutting plan, responding to deformations from tool-tissue interactions. This study examines the integration of physics-based tissue cutting simulations into autonomous robotic surgery to preoperatively predict and compensate for such deformations, aiming to improve surgical precision and reduce the necessity for dynamic adjustments during autonomous surgeries. This study adapts a real-to-sim-to-real workflow. Initially, the Autonomous System for Tumor Resection (ASTR) was employed to evaluate its accuracy in performing preoperatively intended incisions along the irregular contours of porcine tongue pseudotumors. Following this, a finite element analysis-based simulation, utilizing the Simulation Open Framework Architecture (SOFA), was developed and tuned to accurately mimic these tissue and incision interactions. Insights gained from this simulation were applied to refine the robot’s path planning, ensuring a closer alignment of actual incisions with the initially intended surgical plan. The efficacy of this approach was validated by comparing surface incision precision on ex vivo porcine tongues, with the average absolute error reducing from 1.73mm to 1.46mm after applying simulation-driven path adjustments (p < 0.001). Additionally, our method not only demonstrated improvements in maintaining the intended cutting shapes and locations, with shape matching scores using Hu moments enhancing from 0.10 to 0.06 and centroid shifts decreasing from 2.09mm to 1.33mm, but it also potentially reduced the likelihood of adverse oncologic outcomes by preventing clinically suggested excessively close margins of 2.2mm. This feasibility study suggests that merging physics-based cutting simulations with autonomous robotic surgery could potentially lead to more accurate incisions. Show More

Unmanned aerial vehicles (UAVs), known for their agile flight capabilities, require safe trajectory planning to achieve high-speed flights. This is necessary to swiftly evade obstacles and adapt trajectories under hard real-time constraints. These adjustments are essential to generate viable paths that prevent collisions while maintaining high speeds with minimal tracking errors. This paper addresses the challenge of enhancing the safety of agile trajectory planning. The proposed method combines a supervised learning approach, as teacher policy, with deep reinforcement learning (DRL), as student policy. Initially, we train the teacher policy using a path planning algorithm that prioritizes safety while minimizing jerk and flight time. Then, we use this policy to guide the learning of the student policy in various unknown environments. Testing in simulation demonstrates noteworthy advancements, including an 80% reduction in tracking error, a 31% decrease in flight time, a 19% increase in high-speed duration, and a success rate improvement from 50% to 100%, as compared to baseline methods.

Manipulability analysis is a methodology employed to assess the capacity of an articulated system, at a specific configuration, to produce motion or exert force in diverse directions. The conventional method entails generating a virtual ellipsoid using the system’s configuration and model. Yet, this approach poses challenges when applied to systems such as the human body, where direct access to such information is limited, necessitating reliance on estimations. Any inaccuracies in these estimations can distort the ellipsoid’s configuration, potentially compromising the accuracy of the manipulability assessment. To address this issue, this article extends the standard approach by introducing the concept of the manipulability pseudo-ellipsoid. Through a series of theoretical analyses, simulations, and experiments, the article demonstrates that the proposed method exhibits reduced sensitivity to noise in sensory information, consequently enhancing the robustness of the approach.

This study investigates the computational speed and accuracy of two numerical integration methods, cubature and sampling-based, for integrating an integrand over a 2D polygon. Using a group of rovers searching the Martian surface with a limited sensor footprint as a test bed, the relative error and computational time are compared as the area was sub-divided to improve accuracy in the sampling-based approach. The results show that the sampling-based approach exhibits a 14:75% deviation in relative error compared to cubature when it matches the computational performance at 100%. Furthermore, achieving a relative error below 1% necessitates a 10000% increase in relative time to calculate due to the \mathcal{O}\left( {{N^2}} \right)\mathcal{O}\left( {{N^2}} \right) complexity of the sampling-based method. It is concluded that for enhancing reinforcement learning capabilities and other high iteration algorithms, the cubature method is preferred over the sampling-based method.

Environment awareness is crucial for enhancing walking safety and stability of amputee wearing powered prosthesis when crossing uneven terrains such as stairs and obstacles. However, existing environmental perception systems for prosthesis only provide terrain types and corresponding parameters, which fail to prevent potential collisions when crossing uneven terrains and may lead to falls and other severe consequences. In this paper, a visual-inertial motion estimation approach is proposed for prosthesis to perceive its movement and the changes of spatial relationship between the prosthesis and uneven terrain when traversing them. To achieve this, we estimate the knee motion by utilizing a depth camera to perceive the environment and align feature points extracted from uneven terrains. Subsequently, an error-state Kalman filter is incorporated to fuse the inertial data into visual estimations to obtain a more robust and accurate estimation, which is then utilized to derive the motion of the whole prosthesis for our prosthetic control scheme. Experiments conducted on our collected dataset and stair walking trials with powered prosthesis show that the proposed method can accurately track the motion of human leg and the prosthesis with the average root-mean-square error of toe trajectory less than 5 cm. The proposed method is expected to enable the environmental adaptive control for prosthesis, thereby enhancing amputee’s safety and mobility in uneven terrains.

Autonomous driving for urban and highway driving applications often requires High Definition (HD) maps to generate a navigation plan. Nevertheless, various challenges arise when generating and maintaining HD maps at scale. While recent online mapping methods have started to emerge, their performance especially for longer ranges is limited by heavy occlusion in dynamic environments. With these considerations in mind, our work focuses on leveraging lightweight and scalable priors–Standard Definition (SD) maps–in the development of online vectorized HD map representations. We first examine the integration of prototypical rasterized SD map representations into various online mapping architectures. Furthermore, to identify lightweight strategies, we extend the OpenLane-V2 dataset with OpenStreetMaps and evaluate the benefits of graphical SD map representations. A key finding from designing SD map integration components is that SD map encoders are model agnostic and can be quickly adapted to new architectures that utilize bird’s eye view (BEV) encoders. Our results show that making use of SD maps as priors for the online mapping task can significantly speed up convergence and boost the performance of the online centerline perception task by 30% (mAP). Furthermore, we show that the introduction of the SD maps leads to a reduction of the number of parameters in the perception and reasoning task by leveraging SD map graphs while improving the overall performance. Project Page: https://henryzhangzhy.github.io/sdhdmap/.

This paper builds upon the well-established ML-based grasping technique, known as the Grasp-Rectangle (GR) method. The original GR method made two simplifying assumptions: it was designed exclusively for two-finger grippers, and it assumed that the gripper would approach objects solely from a top-down perspective on a horizontal surface. We have extended the GR method, for a multi-finger hand beyond these assumptions to (1) enable grasping from top and side views and (2) engage multiple points of contact, enhancing the algorithm’s overall performance. Our approach leverages geometric cues extracted from object images to calculate the optimal grasp pose and contact points, thereby enhancing grasp reliability. Extensive testing was conducted using a 7DOF robotic arm equipped with a 7-DOF 3-finger gripper. We achieved an accuracy of 98.6% on the Cornell Grasping Dataset with a processing time of 120 milliseconds. Furthermore, when assessing object grasping from both top and side perspectives, our algorithm delivered successful grasps at rates of 95% and 96%, respectively. These findings are rooted in a comprehensive series of tests performed across a diverse array of objects.

Visual object tracking has boosted extensive intelligent applications for unmanned aerial vehicles (UAVs). However, the state-of-the-art (SOTA) enhancers for nighttime UAV tracking always neglect the uneven light distribution in low-light images, inevitably leading to excessive enhancement in scenarios with complex illumination. To address these issues, this work proposes a novel enhancer, i.e., LDEnhancer, enhancing nighttime UAV tracking with light distribution suppression. Specifically, a novel image content refinement module is developed to decompose the light distribution information and image content information in the feature space, allowing for the targeted enhancement of the image content information. Then this work designs a new light distribution generation module to capture light distribution effectively. The features with light distribution information and image content information are fed into the different parameter estimation modules, respectively, for the parameter map prediction. Finally, leveraging two parameter maps, an innovative interweave iteration adjustment is proposed for the collaborative pixel-wise adjustment of low-light images. Additionally, a challenging nighttime UAV tracking dataset with uneven light distribution, namely NAT2024-2, is constructed to provide a comprehensive evaluation, which contains 40 challenging sequences with over 74K frames in total. Experimental results on the authoritative UAV benchmarks and the proposed NAT2024-2 demonstrate that LDEnhancer outperforms other SOTA low-light enhancers for nighttime UAV tracking. Furthermore, real-world tests on a typical UAV platform with an NVIDIA Orin NX confirm the practicality and efficiency of LDEnhancer. The code is available at https: //github.com/vision4robotics/LDEnhancer.

Scanning LiDAR is one of the widely used sensors in autonomous vehicles; however, the inherent sparsity of LiDAR point clouds often affects its performance. To address this issue, upsampling methods could be employed to enhance low-resolution LiDAR data. Although there have been methods on upsampling of single-object point clouds recently in computer vision, they tend to generate a considerable amount of artifacts when dealing with real-world LiDAR scenes consisting of multiple objects. In this paper, we propose a solution to tackle this problem by introducing an instance embedding auxiliary task and a context attention module. With our auxiliary learning architecture, the network can learn features that benefit both the primary upsampling task and the auxiliary instance embedding task. This training design enables the point generation process to be carried out separately and significantly reduces artifacts of the upsampling results on the SemanticKITTI dataset, particularly in areas surrounding instances. By leveraging these techniques to improve the model’s understanding of the relationship between objects and the background in LiDAR scenes, we achieve an overall 4% to 10% improvement in whole-scene upsampling.

To address the leg odometry drift caused by the non-stationary foot contact, this paper introduces a novel data-driven based leg odometry technique for legged robots. By leveraging a Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN), the method learns the biases in the robot’s foot contact locations from sequential IMU measurements and ground reaction forces (GRF). This learned contact bias is then incorporated into the state estimation process using a Kalman filter (KF), significantly improving the precision of leg odometry for legged robots operating in real time. This method, which combines deep learning approaches with conventional filtering techniques, is named the Deep Learning Kalman Filter (DLKF). The effectiveness of the DLKF is demonstrated through simulation and experimental trials using a Unitree Go1 robot across various challenging environments, including uneven terrain, slopes, and stairs, where foot slippage occurs frequently. Our results indicate an average 64.93% reduction in translational errors in leg odometry when the learned contact bias is applied. Further improvements are observed in a fused LiDAR and leg odometry state estimation system, especially in feature-deprived areas, indicating that the proposed leg odometry system can be easily fused with other sensor measurements to get a more precise state estimation.

3D single object tracking (SOT) based on point cloud has attracted much attention due to its important role in machine vision and autonomous driving. Recently, M2-Track proposes a two-stage tracking structure centered on motion, but they ignore the effect of segmentation errors in sparse point cloud scenarios, which hinder the ability of networks to accurately represent tracking targets. To solve the problems, we propose an efficient 3D single object tracker (Abbr. EST) that can effectively segment point cloud features. Firstly, the proposed fusion segmentation module makes up for the feature loss caused by the downsampling strategy and enhances the ability of the network to recognize foreground points. In addition, the global embedded module is used to further focus on the crucial features of the target. This module provides global information by using residual networks and adding background information. Numerous experiments conducted on KITTI and NuScenes benchmarks show that EST achieves superior point cloud tracking in both performance and efficiency.

Robotic assistance is a current research topic with high application value and multiple challenges. Assistive robots are used in various scenarios, such as production lines, operating tables, and elderly care. While providing effective assistance, most of the assistance tasks that current robots can perform are limited to predefined tasks. This limitation arises from the insufficiency of the current robot perception system to forecast future human activities. To address this issue, we propose a novel 2-stage robotic assistant for human activities through future human-object interaction (HOI) segment prediction. Unlike previous work focusing on predefined or short-term tasks, our robotic assistant can make predictions for future assistance according to human habits. In the first stage, we propose a visual-based human-object interaction segment prediction method to predict human activities, which enables the robotic system to infer human intention. Moreover, we define the robotic executable tasks as an interactive tuple to keep the robotic assistance normatively consistent with human activity. Meanwhile, a graph convolutional network with geometric features that can predict human-object interaction segments is proposed to provide target manipulation and target object for the assistive robot. In the second stage, we present a mobile task completion process including visual navigation, object localization and grasping. The perception stage is evaluated on the MPHOI dataset and custom-collected SPHOI dataset. Finally, we evaluate our comprehensive framework through real-time experimentation.

This paper introduces an innovation of the Spherical Omnidirectional Wheel, designed to achieve omnidirectional driving motion. In previous models, the supporting shaft was placed at the center of the mechanism. However, achieving both smoothness and high load-capacity in such designs proved challenging. The mechanism proposed in this study features an offset design, enabling outer support for the wheel. A prototype was developed and its basic motion was experimentally validated.