ICRA 2021 Papers

Neuromorphic processors like Loihi offer a promising alternative to conventional computing modules for endowing constrained systems like micro air vehicles (MAVs) with robust, efficient and autonomous skills such as take-off and landing, obstacle avoidance, and pursuit. However, a major challenge for using such processors on robotic platforms is the reality gap between simulation and the real world. In this study, we present for the very first time a fully embedded application of the Loihi neuromorphic chip prototype in a flying robot. A spiking neural network (SNN) was evolved to compute the thrust command based on the divergence of the ventral optic flow field to perform autonomous landing. Evolution was performed in a Python-based simulator using the PySNN library. The resulting network architecture consists of only 35 neurons distributed among 3 layers. Quantitative analysis between simulation and Loihi reveals a root-mean-square error of the thrust setpoint as low as 0.005 g, along with a 99.8% matching of the spike sequences in the hidden layer, and 99.7% in the output layer. The proposed approach successfully bridges the reality gap, offering important insights for future neuromorphic applications in robotics. Supplementary material is available at https://mavlab.tudelft.nl/loihi/.

Differentiable simulators provide an avenue for closing the sim-to-real gap by enabling the use of efficient, gradient-based optimization algorithms to find the simulation parameters that best fit the observed sensor readings. Nonetheless, these analytical models can only predict the dynamical behavior of systems for which they have been designed. In this work, we study the augmentation of a novel differentiable rigid-body physics engine via neural networks that is able to learn nonlinear relationships between dynamic quantities and can thus model effects not accounted for in traditional simulators. Such augmentations require less data to train and generalize better compared to entirely data-driven models. Through extensive experiments, we demonstrate the ability of our hybrid simulator to learn complex dynamics involving frictional contacts from real data, as well as match known models of viscous friction, and present an approach for automatically discovering useful augmentations. We show that, besides benefiting dynamics modeling, inserting neural networks can accelerate model-based control architectures. We observe a ten-fold speedup when replacing the QP solver inside a model-predictive gait controller for quadruped robots with a neural network, allowing us to significantly improve control delays as we demonstrate in real-hardware experiments. We publish code, additional results and videos from our experiments on our project webpage at https://sites.google.com/usc.edu/neuralsim.

We consider the problem of optimizing a robot morphology to achieve the best performance for a target task, under computational resource limitations. The evaluation process for each morphological design involves learning a controller for the design, which can consume substantial time and computational resources. To address the challenge of expensive robot morphology evaluation, we present a continuous multi-fidelity Bayesian Optimization framework that efficiently utilizes computational resources via low-fidelity evaluations. We identify the problem of non-stationarity over fidelity space. Our proposed fidelity warping mechanism can learn representations of learning epochs and tasks to model non-stationary covariances between continuous fidelity evaluations which prove challenging for off-the-shelf stationary kernels. Various experiments demonstrate that our method can utilize the low-fidelity evaluations to efficiently search for the optimal robot morphology, outperforming state-of-the-art methods.

We have recently proposed two pile loading controllers that learn from human demonstrations: a neural network (NNet) [1] and a random forest (RF) controller [2]. In the field experiments the RF controller obtained clearly better success rates. In this work, the previous findings are drastically revised by experimenting summer time trained controllers in winter conditions. The winter experiments revealed a need for additional sensors, more training data, and a controller that can take advantage of these. Therefore, we propose a revised neural controller (NNetV2) which has a more expressive structure and uses a neural attention mechanism to focus on important parts of the sensor and control signals. Using the same data and sensors to train and test the three controllers, NNetV2 achieves better robustness against drastically changing conditions and superior success rate. To the best of our knowledge, this is the first work testing a learning-based controller for a heavy-duty machine in drastically varying outdoor conditions and delivering high success rate in winter, being trained in summer.

Uncertainty estimation for point cloud semantic segmentation is to quantify the confidence degree for the predicted label of points, which is essential for decision-making tasks. This paper proposes a neighborhood spatial aggregation based method, NSA-MC dropout, to achieve efficient uncertainty estimation for point cloud semantic segmentation. Unlike the traditional uncertainty estimation method MC dropout de-pending on repeated inferences, our NSA-MC dropout achieves uncertainty estimation through one-time inference. Specifically, a space-dependent method is designed to sample the model many times by performing stochastic forward pass through the model just once, and it approximates the repeated inferences based sampling process in MC dropout. Besides, a neighborhood spatial aggregation module, called NSA, aggregates neighborhood probabilistic outputs for each point and works with space-dependent sampling to establish output distribution. Finally, we propose an uncertainty-aware framework NSA-MC dropout to capture the uncertainty of prediction results efficiently. Experimental results show that our method obtains comparable performance with MC dropout. More significantly, our NSA-MC dropout has little influence on the efficiency of semantic inference. It is much faster than MC dropout, and the inference time does not establish a coupling relation with the sampling times. Our code is available at https://github.com/chaoqi7/Uncertainty_Estimation_PCSS

An underlying structure in several sampling-based methods for continuous multi-robot motion planning (MRMP) is the tensor roadmap (PR), which emerges from combining multiple PRM graphs constructed for the individual robots via a tensor product. We study the conditions under which the TR encodes a near-optimal solution for MRMP—satisfying these conditions implies near optimality for a variety of popular planners, including dRRT*, and the discrete methods M* and CBS when applied to the continuous domain. We develop the first finite-sample analysis of this kind, which specifies the number of samples, their deterministic distribution, and magnitude of the connection radii that should be used by each individual PRM graph, to guarantee near-optimality using the TR. This significantly improves upon a previous asymptotic analysis, wherein the number of samples tends to infinity. Our new finite sample-size analysis supports guaranteed high- quality solutions in practice within finite time. To achieve our new result, we first develop a sampling scheme, which we call the staggered grid, for finite-sample motion planning for individual robots, which requires significantly less samples than previous work. We then extend it to the much more involved MRMP setting which requires to account for interactions among multiple robots. Finally, we report on a few experiments that serve as a verification of our theoretical findings and raise interesting questions for further investigation.

Robot navigation is a task where reinforcement learning approaches are still unable to compete with traditional path planning. State-of-the-art methods differ in small ways, and do not all provide reproducible, openly available implementations. This makes comparing methods a challenge. Recent research has shown that unsupervised learning methods can scale impressively, and be leveraged to solve difficult problems. In this work, we design ways in which unsupervised learning can be used to assist reinforcement learning for robot navigation. We train two end-to-end, and 18 unsupervised-learning-based architectures, and compare them, along with existing approaches, in unseen test cases. We demonstrate our approach working on a real life robot. Our results show that unsupervised learning methods are competitive with end-to-end methods. We also highlight the importance of various components such as input representation, predictive unsupervised learning, and latent features. We make all our models publicly available, as well as training and testing environments, and tools 1. This release also includes OpenAI-gym-compatible environments designed to emulate the training conditions described by other papers, with as much fidelity as possible. Our hope is that this helps in bringing together the field of RL for robot navigation, and allows meaningful comparisons across state-of-the-art methods.

This paper presents a novel non-Gaussian inference algorithm, Normalizing Flow iSAM (NF-iSAM), for solving SLAM problems with non-Gaussian factors and/or non-linear measurement models. NF-iSAM exploits the expressive power of neural networks, and trains normalizing flows to draw samples from the joint posterior of non-Gaussian factor graphs. By leveraging the Bayes tree, NF-iSAM is able to exploit the sparsity structure of SLAM, thus enabling efficient incremental updates similar to iSAM2, albeit in the more challenging non- Gaussian setting. We demonstrate the performance of NF-iSAM and compare it against the state-of-the-art algorithms such as iSAM2 (Gaussian) and mm-iSAM (non-Gaussian) in synthetic and real range-only SLAM datasets.

3D point cloud-based place recognition is highly demanded by autonomous driving in GPS-challenged environments and serves as an essential component (i.e. loop-closure detection) in lidar-based SLAM systems. This paper proposes a novel approach, named NDT-Transformer, for real-time and large-scale place recognition using 3D point clouds. Specifically, a 3D Normal Distribution Transform (NDT) representation is employed to condense the raw, dense 3D point cloud as probabilistic distributions (NDT cells) to provide the geometrical shape description. Then a novel NDT-Transformer network learns a global descriptor from a set of 3D NDT cell representations. Benefiting from the NDT representation and NDT-Transformer network, the learned global descriptors are enriched with both geometrical and contextual information. Finally, descriptor retrieval is achieved using a query-database for place recognition. Compared to the state-of-the-art methods, the proposed approach achieves an improvement of 7.52% on average top 1 recall and 2.73% on average top 1% recall on the Oxford Robotcar benchmark.

Unmanned aerial vehicle (UAV) based visual tracking has been confronted with numerous challenges, e.g., object motion and occlusion. These challenges generally introduce unexpected mutations of target appearance and result in tracking failure. However, prevalent discriminative correlation filter (DCF) based trackers are insensitive to target mutations due to a predefined label, which concentrates on merely the centre of the training region. Meanwhile, appearance mutations caused by occlusion or similar objects usually lead to the inevitable learning of wrong information. To cope with appearance mutations, this paper proposes a novel DCF-based method to enhance the sensitivity and resistance to mutations with an adaptive hybrid label, i.e., MSCF. The ideal label is optimized jointly with the correlation filter and remains temporal consistency. Besides, a novel measurement of mutations called mutation threat factor (MTF) is applied to correct the label dynamically. Considerable experiments are conducted on widely used UAV benchmarks. The results indicate that the performance of MSCF tracker surpasses other 26 state-of-the- art DCF-based and deep-based trackers. With a real-time speed of ~38 frames/s, the proposed approach is sufficient for UAV tracking commissions.

The development and control of biological actuators have been an active research field. Biological actuators revealed high mobility with compact dimensions, which is critical for the design of microrobots. The powerful kicking motion of the locust is important for its quick jumping. Herein, we examined the kicking process of the locust’s hindleg and controlled the flexion and extension motions via exogenous stimulation. Unlike a simple extension of the leg, co-contraction is adopted by locust to store energy and increase jumping power. Thus, we imitated the co-contraction process and transformed the locust’s hindleg into a biological jumping actuator. Through coordinating the kicking of bilateral hindlegs together, we achieved the jumping control of a locust.

To address the problem of building accurate and coherent models for large-scale terrains from incomplete and noisy sensor data, this paper proposes a novel framework that can efficiently infer terrain structures by divisionally providing the best linear unbiased estimates for the elevation values. To avoid data ambiguity caused by the uncertainty of sensor data, the proposed method introduces elevation filtering to extract the terrain surfaces, which reduces the amount of data greatly while the contained terrain information is basically unchanged. Then, for the large-scale terrains, the Gaussian mixture model is used to divide the interested regions, which remarkably improves the prediction accuracy and speed. Finally, for each subregion, a gaussian process regression model based on the static kernel is used to create a multiresolution terrain representation, which can deal with incompleteness of sensor data by considering the spatial correlations of the terrain. Evaluations of the proposed technique were conducted on diverse large-scale field terrains, including the quarry, planetary emulation terrain and highland, showing that the proposed method outperforms the state-of-art terrain modeling techniques in terms of the prediction accuracy, computation speed and memory consumption. As a practical application, the path planning problem was explored based on this terrain modeling technique to produce a better path.

Scores of papers show, given some robots, how to improve the useful work they perform. Continuing this line, we consider the efficiency of robot experiments by examining the feasibility of conducting several experiments simultaneously, interleaving execution and sharing resources between them. This paper lays theoretical groundwork for that concept and demonstrates its feasibility and utility.

The goal of foraging robot swarms is to search and deliver resources to a specific central collection zone quickly. In the previously proposed multiple-place foraging algorithm with dynamic depots, foraging performance decreases as search areas and swarm sizes increase: depots need to travel long distances to deliver resources to the center, and more robots produce more congestion on their journeys. We propose a novel extension to the multiple-place foraging in which multiple robot chains are deployed dynamically. Each robot chain connects a foraging location to the central collection zone. Instead of delivering resources by a single robot, resources are passed on robot chains from foraging locations to the center directly such that congestion near the central collection zone can be avoided. Dynamic robot chains can also relocate themselves to get closer to the resources while avoiding obstacles. We simulate our robot swarms in the robot simulator ARGoS. Our experiments show that robots using the MPFA with dynamic chains outperform the MPFA with dynamic depots and have less congestion.

Dynamics modeling in outdoor and unstructured environments is difficult because different elements in the environment interact with the robot in ways that can be hard to predict. Leveraging multiple sensors to perceive maximal information about the robot’s environment is thus crucial when building a model to perform predictions about the robot’s dynamics with the goal of doing motion planning. We design a model capable of long-horizon motion predictions, leveraging vision, lidar and proprioception, which is robust to arbitrarily missing modalities at test time. We demonstrate in simulation that our model is able to leverage vision to predict traction changes. We then test our model using a real-world challenging dataset of a robot navigating through a forest, performing predictions in trajectories unseen during training. We try different modality combinations at test time and show that, while our model performs best when all modalities are present, it is still able to perform better than the baseline even when receiving only raw vision input and no proprioception, as well as when only receiving proprioception. Overall, our study demonstrates the importance of leveraging multiple sensors when doing dynamics modeling in outdoor conditions.

Dense depth estimation is essential to scene-understanding for autonomous driving. However, recent self-supervised approaches on monocular videos suffer from scale-inconsistency across long sequences. Utilizing data from the ubiquitously copresent global positioning systems (GPS), we tackle this challenge by proposing a dynamically-weighted GPS-to-Scale (g2s) loss to complement the appearance-based losses. We emphasize that the GPS is needed only during the multimodal training, and not at inference. The relative distance between frames captured through the GPS provides a scale signal that is independent of the camera setup and scene distribution, resulting in richer learned feature representations. Through extensive evaluation on multiple datasets, we demonstrate scale-consistent and -aware depth estimation during inference, improving the performance even when training with low-frequency GPS data.

Object slip perception is essential for mobile manipulation robots to perform manipulation tasks reliably in the dynamic real-world. Traditional approaches to robot arms’ slip perception use tactile or vision sensors. However, mobile robots still have to deal with noise in their sensor signals caused by the robot’s movement in a changing environment. To solve this problem, we present an anomaly detection method that utilizes multisensory data based on a deep autoencoder model. The proposed framework integrates heterogeneous data streams collected from various robot sensors, including RGB and depth cameras, a microphone, and a force-torque sensor. The integrated data is used to train a deep autoencoder to construct latent representations of the multisensory data that indicate the normal status. Anomalies can then be identified by error scores measured by the difference between the trained encoder’s latent values and the latent values of reconstructed input data. In order to evaluate the proposed framework, we conducted an experiment that mimics an object slip by a mobile service robot operating in a real-world environment with diverse household objects and different moving patterns. The experimental results verified that the proposed framework reliably detects anomalies in object slip situations despite various object types and robot behaviors, and visual and auditory noise in the environment.

Within the Multi-Platform Inspection, Maintenance and Repair in Extreme Environments (MIMRee) project, a lightweight and multifunctional robotic repair arm is created for wind turbine blades. The design features a toolbox at the base of the arm housing multiple end-effector tools and an autonomous end-effector tool-changer. The arm communicates commands and data via internet with a bespoke user interface enabling human-in-the-loop operation and overriding of autonomous repair actions. This paper outlines our approach in design, development, testing and control of the robotic repair system. The functionalities of the arm include cleaning, sanding, and filler material deposition and forming, each using a bespoke end-effector tool closely replicating the relevant manual repair process. The experimental results confirm the effectiveness of our approach indicating a maximum end-effector position error of 3 mm, a maximum tool switching time of 8 seconds, and a maximum arm’s weight of 1.8 kg. This presents around 84% weight reduction compared with existing technologies used for the same purpose. Our standalone design enables modular integration into a wide range of mobile platform types used in industrial operations.

As a common cause of anemia and gastrointestinal bleeding, angiodysplasia (AD) diagnosis in wireless capsule endoscopy (WCE) images is important in clinical. Current manual review requires undivided concentration of the gastroenterologists, which is laborious and time-consuming. The development of computational methods that can assist automated diagnosis of angiodysplasia is highly desirable. In this paper, we present a new approach, ADNet, for angiodysplasia segmentation using convolutional neural networks (CNNs). Compared with previous learning strategies, ADNet gains accuracy from attentionguided and domain-adversarial training via a multibranch CNN architecture. Specifically, the core branch is constructed for AD segmentation in a fully convolutional manner. Then we propose an attention module embedded in the attention branch to enhance network feature learning, which allows ADNet to focus on the most informative and AD relevant regions while processing. Furthermore, an adaptation branch is built to learn domain-invariant features by adversarial training, aiming to improve the performance when datasets are expanded while preventing the degradation induced by the variations in WCE image acquisition. ADNet is evaluated using two WCE datasets with angiodysplasia and the results show the accuracy gains we obtain, where the state-of-the-art segmentation performance on the public dataset of GIANA’17 is achieved.