ICRA 2023 Papers

General scene understanding for robotics requires flexible semantic representation, so that novel objects and structures which may not have been known at training time can be identified, segmented and grouped. We present an algorithm which fuses general learned features from a standard pre-trained network into a highly efficient 3D geometric neural field representation during real-time SLAM. The fused 3D feature maps inherit the coherence of the neural field's geometry representation. This means that tiny amounts of human labelling interacting at runtime enable objects or even parts of objects to be robustly and accurately segmented in an open set manner. Project page: https://makezur.github.io/FeatureRealisticFusion/

Deep reinforcement learning (DRL) methods can solve complex continuous control tasks in simulated environments by taking actions based solely on state observations at each decision point. Because of the dynamics involved, individual snapshots of real-world sensor measurements afford only partial state observability, so it is typical to use a history of observations to improve training and policy performance. Such intertemporal information can be further exploited using a recurrent neural network (RNN) to reduce the dimensionality of the dynamic state representation. However, using RNNs as an internal part of a DRL network presents challenges of its own; and even then, the improvements in resulting policies are usually limited. To address these shortcomings, we propose using gated feature extraction to improve DRL training of real-world robots. Specifically, we use an untrained gated recurrent unit (GRU) to encode a low-dimension representation of the state observation sequence before passing it to the DRL training procedure. In addition to dimensionality reduction, this allows us to unroll the RNN by encoding the observations cumulatively as they are collected, thereby avoiding same-length input requirements, and train the RL network on the raw observations at the current step combined with the GRU-encoding of the preceding steps. Our simulation experiments employ gated feature extraction with the TD3 algorithm. Our results show that the GRU-encoded state observations improve the training speed and execution performance of the TD3 algorithm, improving the learned policies in all 19 test cases, exceeding the maximum achieved reward by over 38% in 8 and doubling the maximum achieved reward in three, while also outperforming a baseline implementation of SAC in 17 out of 19 environments. Moreover, the greatest improvement is seen in real-world experiments, where our approach successfully learns to balance a pendulum as well as a complex quadrupedal locomotion task. In c... Show More

Multi-finger grasping relies on high quality training data, which is hard to obtain: human data is hard to transfer and synthetic data relies on simplifying assumptions that reduce grasp quality. By making grasp simulation differentiable, and contact dynamics amenable to gradient-based optimization, we accelerate the search for high-quality grasps with fewer limiting assumptions. We present Grasp'D-1M: a large-scale dataset for multi-finger robotic grasping, synthesized with Fast-Grasp'D, a novel differentiable grasping simulator. Grasp'D-1M contains one million training examples for three robotic hands (three, four and five-fingered), each with multimodal visual inputs (RGB+depth+segmentation, available in mono and stereo). Grasp synthesis with Fast-Grasp'D is 10x faster than GraspIt! [1] and 20x faster than the prior Grasp'D differentiable simulator [2]. Generated grasps are more stable and contact-rich than GraspIt! grasps, regardless of the distance threshold used for contact generation. We validate the usefulness of our dataset by retraining an existing vision-based grasping pipeline [3] on Grasp'D-1M, and showing a dramatic increase in model performance, predicting grasps with 30% more contact, a 33% higher epsilon metric, and 35% lower simulated displacement. Additional details at fast-graspd.github.io.

This paper presents the design and performance of a planar 3R robot capable of dexterous constrained manipulation when interacting with a stiff environment. A novel variable stiffness actuator (VSA) having a stiffness ratio of approximately 500 is also described. Variable stiffness actuation, together with a combined position/compliance manipulation path, is used to: 1) allow the robot to passively comply with its environment along kinematically constrained directions despite model error in constraint locations, and 2) generate high stiffness for accurate motion control along kinematically unconstrained directions despite resisting forces. This manipulation strategy provides dexterity for cases in which mechanical work must be performed while complying with constraints. The manipulation strategy and robot performance were evaluated with the task of turning a steel crank to lift a weight. Results show that, when using passive compliance control, the robot completed the task 29 times faster with constraint forces 80% lower than when using traditional active compliance control (with VSAs at their highest stiffness).

In active source seeking, a robot takes repeated measurements in order to locate a signal source in a cluttered and unknown environment. A key component of an active source seeking robot planner is a model that can produce estimates of the signal at unknown locations with uncertainty quantification. This model allows the robot to plan for future measurements in the environment. Traditionally, this model has been in the form of a Gaussian process, which has difficulty scaling and cannot represent obstacles. We propose a global and local factor graph model for active source seeking, which allows the model to scale to a large number of measurements and represent unknown obstacles in the environment. We combine this model with extensions to a highly scalable planner to form a system for large-scale active source seeking. We demonstrate that our approach outperforms baseline methods in both simulated and real robot experiments.

Enlightened by the fast-running gait of mammals like cheetahs and wolves, we design and fabricate a single-actuated untethered compliant robot that is capable of galloping at a speed of 313 mm/s or 1.56 body length per second (BL/s), faster than most reported soft crawlers in mm/s and BL/s. An in-plane prestressed hair clip mechanism (HCM) made up of semirigid materials, i.e. plastics are used as the supporting chassis, the compliant spine, and the force amplifier of the robot at the same time, enabling the robot to be simple, rapid, and strong. With experiments, we find that the HCM robotic locomotion speed is linearly related to actuation frequencies and substrate friction differences except for concrete surface, that tethering slows down the crawler, and that asymmetric actuation creates a new galloping gait. This paper demonstrates the potential of HCM-based soft robots.

Robotic systems need advanced mobility capabili-ties to operate in complex, three-dimensional environments designed for human use, e.g., multi-level buildings. Incorporating some level of autonomy enables robots to operate robustly, reliably, and efficiently in such complex environments, e.g., automatically “returning home” if communication between an operator and robot is lost during deployment. This work presents a novel method that enables mobile robots to robustly operate in multi-level environments by making it possible to autonomously locate and climb a range of different staircases. We present results wherein a wheeled robot works together with a quadrupedal system to quickly detect different staircases and reliably climb them. The performance of this novel staircase detection algorithm that is able to run on the heterogeneous platforms is compared to the current state-of-the-art detection algorithm. We show that our approach significantly increases the accuracy and speed at which detections occur.

Unmanned aerial vehicles assist in maritime search and rescue missions by flying over large search areas to autonomously search for objects or people. Reliably detecting objects of interest requires fast models to employ on embedded hardware. Moreover, with increasing distance to the ground station only part of the video data can be transmitted. In this work, we consider the problem of finding meaningful region of interest proposals in a video stream on an embedded GPU. Current object or anomaly detectors are not suitable due to their slow speed, especially on limited hardware and for large image resolutions. Lastly, objects of interest, such as pieces of wreckage, are often not known a priori. Therefore, we propose an end-to-end future frame prediction model running in real-time on embedded GPUs to generate region proposals. We analyze its performance on large-scale maritime data sets and demonstrate its benefits over traditional and modern methods.

In this paper, we propose a fast extrinsic calibration method for fusing multiple inertial measurement units (MIMU) to improve visual-inertial odometry (VIO) localization accuracy. Currently, data fusion algorithms for MIMU highly depend on the number of inertial sensors. Based on the assumption that extrinsic parameters between inertial sensors are perfectly calibrated, the fusion algorithm provides better localization accuracy with more IMUs, while neglecting the effect of extrinsic calibration error. Our method builds two non-linear least-squares problems to estimate the MIMU relative position and orientation separately, independent of external sensors and inertial noises online estimation. Then we give the general form of the virtual IMU (VIMU) method and propose its propagation on manifold. We perform our method on datasets, our self-made sensor board, and board with different IMUs, validating the superiority of our method over competing methods concerning speed, accuracy, and robustness. In the simulation experiment, we show that only fusing two IMUs with our calibration method to predict motion can rival nine IMUs. Real-world experiments demonstrate better localization accuracy of the VIO integrated with our calibration method and VIMU propagation on manifold.

Motion deblurring is a critical ill-posed problem that is important in many vision-based robotics applications. The recently proposed event-based double integral (EDI) provides a theoretical framework for solving the deblurring prob-lem with the event camera and generating clear images at high frame-rate. However, the original EDI is mainly designed for offline computation and does not support real-time requirement in many robotics applications. In this paper, we propose the fast EDI, an efficient implementation of EDI that can achieve real-time online computation on single-core CPU devices, which is common for physical robotic platforms used in practice. In experiments, our method can handle event rates at as high as 13 million event per second in a wide variety of challenging lighting conditions. We demonstrate the benefit on multiple downstream real-time applications, including localization, vi-sual tag detection, and feature matching.

Self-assessment rules play an essential role in safe and effective real-world robotic applications, which verify the feasibility of the selected action before actual execution. But how to utilize the self-assessment results to re-choose actions remains a challenge. Previous methods eliminate the selected action evaluated as failed by the self-assessment rules, and re-choose one with the next-highest affordance (i.e. process-of-elimination strategy [1]), which ignores the dependency between the self-assessment results and the remaining untried actions. However, this dependency is important since the previous failures might help trim the remaining over-estimated actions. In this paper, we set to investigate this dependency by learning a failure-aware policy. We propose two architectures for the failure-aware policy by representing the self-assessment results of previous failures as the variable state, and leveraging recurrent neural networks to implicitly memorize the previous failures. Experiments conducted on three tasks demonstrate that our method can achieve better performances with higher task success rates by less trials. Moreover, when the actions are correlated, learning a failure-aware policy can achieve better performance than the process-of-elimination strategy.

Motion prediction is essential for safe and efficient autonomous driving. However, the inexplicability and uncertainty of complex artificial intelligence models may lead to unpredictable failures of the motion prediction module, which may mislead the system to make unsafe decisions. Therefore, it is necessary to develop methods to guarantee reliable autonomous driving, where failure detection is a potential direction. Uncertainty estimates can be used to quantify the degree of confidence a model has in its predictions and may be valuable for failure detection. We propose a framework of failure detection for motion prediction from the uncertainty perspective, considering both motion uncertainty and model uncertainty, and formulate various uncertainty scores according to different prediction stages. The proposed approach is evaluated based on different motion prediction algorithms, uncertainty estimation methods, uncertainty scores, etc., and the results show that uncertainty is promising for failure detection for motion prediction but should be used with caution.

Failure detection and fault tolerant control are fundamental safety features of any aerial vehicle. With the emer-gence of complex, multi-body flying systems such as jet-powered humanoid robots, it becomes of crucial importance to design fault detection and control strategies for these systems, too. In this paper we propose a fault detection and control framework for the flying humanoid robot iRonCub in case of loss of one turbine. The framework is composed of a failure detector based on turbines rotational speed, a momentum-based flight control for fault response, and an offline reference generator that produces far-from-singularities configurations and accounts for self and jet exhausts collision avoidance. Simulation results with Gazebo and MATLAB prove the effectiveness of the proposed control strategy.

Accurate localization is a core component of a robot's navigation system. To this end, global navigation satellite systems (GNSS) can provide absolute measurements outdoors and, therefore, eliminate long-term drift. However, fusing GNSS data with other sensor data is not trivial, especially when a robot moves between areas with and without sky view. We propose a robust approach that tightly fuses raw GNSS receiver data with inertial measurements and, optionally, lidar observations for precise and smooth mobile robot localization. A factor graph with two types of GNSS factors is proposed. First, factors based on pseudoranges, which allow for global localization on Earth. Second, factors based on carrier phases, which enable highly accurate relative localization, which is useful when other sensing modalities are challenged. Unlike traditional differential GNSS, this approach does not require a connection to a base station. On a public urban driving dataset, our approach achieves accuracy comparable to a state-of-the-art algorithm that fuses visual inertial odometry with GNSS data-despite our approach not using the camera, just inertial and GNSS data. We also demonstrate the robustness of our approach using data from a car and a quadruped robot moving in environments with little sky visibility, such as a forest. The accuracy in the global Earth frame is still 1–2 m, while the estimated trajectories are discontinuity-free and smooth. We also show how lidar measurements can be tightly integrated. We believe this is the first system that fuses raw GNSS observations (as opposed to fixes) with lidar in a factor graph.

Robot grasping has been widely studied in the last decade. Recently, Deep Learning made possible to achieve remarkable results in grasp pose estimation, using depth and RGB images. However, only few works consider the choice of the object to grasp. Moreover, they require a huge amount of data for generalizing to unseen object categories. For this reason, we introduce the Few-shot Semantic Grasping task where the objective is inferring a correct grasp given only five labelled images of a target unseen object. We propose a new deep learning architecture able to solve the aforementioned problem, leveraging on a Few-shot Semantic Segmentation module. We have evaluated the proposed model both in the Graspnet dataset and in a real scenario. In Graspnet, we achieve 40,95% accuracy in the Few-shot Semantic Grasping task, outperforming baseline approaches. In the real experiments, the results confirmed the generalization ability of the network.

Painting is an artistic process of rendering visual content that achieves the high-level communication goals of an artist that may change dynamically throughout the creative process. In this paper, we present a Framework and Robotics Initiative for Developing Arts (FRIDA) that enables humans to produce paintings on canvases by collaborating with a painter robot using simple inputs such as language descriptions or images. FRIDA introduces several technical innovations for computationally modeling a creative painting process. First, we develop a fully differentiable simulation environment for painting, adopting the idea of real to simulation to real (real2sim2real). We show that our proposed simulated painting environment is higher fidelity to reality than existing simulation environments used for robot painting. Second, to model the evolving dynamics of a creative process, we develop a planning approach that can continuously optimize the painting plan based on the evolving canvas with respect to the high-level goals. In contrast to existing approaches where the content generation process and action planning are performed independently and sequentially, FRIDA adapts to the stochastic nature of using paint and a brush by continually re-planning and re-assessing its semantic goals based on its visual perception of the painting progress. We describe the details on the technical approach as well as the system integration. FRIDA software is freely available at: https://github.com/cmubig/Frida.

This article presents a 3D point cloud map-merging framework for egocentric heterogeneous multi-robot exploration, based on overlap detection and alignment, that is independent of a manual initial guess or prior knowledge of the robots' poses. The novel proposed solution utilizes state-of-the-art place recognition learned descriptors, that through the framework's main pipeline, offer a fast and robust region overlap estimation, hence eliminating the need for the time-consuming global feature extraction and feature matching process that is typically used in 3D map integration. The region overlap estimation provides a homogeneous rigid transform that is applied as an initial condition in the point cloud registration algorithm Fast-GICP, which provides the final and refined alignment. The efficacy of the proposed framework is experimentally evaluated based on multiple field multi-robot exploration missions in underground environments, where both ground and aerial robots are deployed, with different sensor configurations.

Federated learning is a promising technique for training global models in a data-decentralized environment. In this paper, we propose a federated learning approach for robotic object grasping. The main challenge is that the data collected by multiple robots deployed in different environments tends to form heterogeneous data distributions (i.e., non-IID) and that the existing federated learning methods on such data distributions show serious performance degradation. To tackle this problem, we propose federated object grasping learning (FOGL) that uses cross-evaluation in a general federated learning process to assess the training performance of robots. We cluster robots with similar training patterns and perform independent federated learning on each cluster. Finally, we integrate the global models for each cluster through an ensemble inference. We apply FOGL to various federated learning scenarios in robotic object grasping and show state-of-the-art performance on the Cornell grasping dataset.

It has become a consensus that autonomous vehicles (AVs) will first be widely deployed on highways. However, the complexity of highway interchanges becomes the bottleneck for their deployment. An AV should be sufficiently tested under different highway interchanges, which is still challenging due to the lack of available datasets containing diverse highway interchanges. In this paper, we propose a model-driven method, Flyover, to generate a dataset of diverse interchanges with measurable diversity coverage. First, Flyover uses a labeled digraph to model interchange topology. Second, Flyover takes real-world interchanges as input to guarantee topology practicality and extracts different topology equivalence classes by classifying corresponding topology models. Third, for each topology class, Flyover identifies the corresponding geometrical features for the ramps and generates concrete interchanges using k-way combinatorial coverage and differential evolution. To illustrate the diversity and applicability of the generated interchange dataset, we test the built-in traffic flow control algorithm in SUMO and the fuel-optimization trajectory tracking algorithm deployed to Alibaba's autonomous trucks on the dataset. The results show that except for the geometrical difference, the interchanges are diverse in throughput and fuel consumption under the traffic flow control and trajectory tracking algorithms, respectively.