ICRA 2021 Papers

Model-based controllers on real robots require accurate knowledge of the system dynamics to perform optimally. For complex dynamics, first-principles modeling is not sufficiently precise, and data-driven approaches can be leveraged to learn a statistical model from real experiments. However, the efficient and effective data collection for such a data-driven system on real robots is still an open challenge. This paper introduces an optimization problem formulation to find an informative trajectory that allows for efficient data collection and model learning. We present a sampling-based method that computes an approximation of the trajectory that minimizes the prediction uncertainty of the dynamics model. This trajectory is then executed, collecting the data to update the learned model. We experimentally demonstrate the capabilities of our proposed framework when applied to a complex omnidirectional flying vehicle with tiltable rotors. Using our informative trajectories results in models which outperform models obtained from non-informative trajectory by 13.3% with the same amount of training data. Furthermore, we show that the model learned from informative trajectories generalizes better than the one learned from non-informative trajectories, achieving better tracking performance on different tasks.

Trajectory optimization of sensing robots to actively gather information of targets has received much attention in the past. It is well-known that under the assumption of linear Gaussian target dynamics and sensor models the stochastic Active Information Acquisition problem is equivalent to a deterministic optimal control problem. However, the above-mentioned assumptions regarding the target dynamic model are limiting. In real-world scenarios, the target may be subject to disturbances whose models or statistical properties are hard or impossible to obtain. Typical scenarios include abrupt maneuvers, jumping disturbances due to interactions with the environment, anomalous misbehaviors due to system faults/attacks, etc. Motivated by the above considerations, in this paper we consider targets whose dynamic models are subject to arbitrary unknown inputs whose models or statistical properties are not assumed to be available. In particular, with the aid of an unknown input decoupled filter, we formulate the sensor trajectory planning problem to track evolution of the target state and analyse the resulting performance for both the state and unknown input evolution tracking. Inspired by concepts of Reduced Value Iteration, a suboptimal solution that expands a search tree via Forward Value Iteration with informativeness-based pruning is proposed. Concrete suboptimality performance guarantees for tracking both the state and the unknown input are established. Numerical simulations of a target tracking example are presented to compare the proposed solution with a greedy policy.

This work presents an approach for control, state-estimation and learning model (hyper)parameters for robotic manipulators. It is based on the active inference framework, prominent in computational neuroscience as a theory of the brain, where behaviour arises from minimizing variational free-energy. First, we show there is a direct relationship between active inference controllers, and classic methods such as PID control. We demonstrate its application for adaptive and robust behaviour of a robotic manipulator that rivals state-of-the-art. Additionally, we show that by learning specific hyperparameters, our approach can deal with unmodeled dynamics, damps oscillations, and is robust against poor initial parameters. The approach is validated on the ‘Franka Emika Panda’ 7 DoF manipulator. Finally, we highlight limitations of active inference controllers for robotic systems.

Many robot applications call for autonomous exploration and mapping of unknown and unstructured environments. Information-based exploration techniques, such as Cauchy-Schwarz quadratic mutual information (CSQMI) and fast Shannon mutual information (FSMI), have successfully achieved active binary occupancy mapping with range measurements. However, as we envision robots performing complex tasks specified with semantically meaningful objects, it is necessary to capture semantic categories in the measurements, map representation, and exploration objective. This work develops a Bayesian multi-class mapping algorithm utilizing range-category measurements. We derive a closed-form efficiently computable lower bound for the Shannon mutual information between the multi-class map and the measurements. The bound allows rapid evaluation of many potential robot trajectories for autonomous exploration and mapping. We compare our method against frontier-based and FSMI exploration and apply it in a 3-D photo-realistic simulation environment.

Modular soft robots combine the strengths of two traditionally separate areas of robotics. As modular robots, they can show robustness to individual failure and reconfigurability; as soft robots, they can deform and undergo large shape changes in order to adapt to their environment, and have inherent human safety. However, for sensing and communication these robots also combine the challenges of both: they require solutions that are scalable (low cost and complexity) and efficient (low power) to enable collectives of large numbers of robots, and these solutions must also be able to interface with the high extension ratio elastic bodies of soft robots. In this work, we seek to address these challenges using acoustic signals produced by piezoelectric surface transducers that are cheap, simple, and low power, and that not only integrate with but also leverage the elastic robot skins for signal transmission. Importantly, to further increase scalability, the transducers exhibit multi-functionality made possible by a relatively flat frequency response across the audible and ultrasonic ranges. With minimal hardware, they enable directional contact-based communication, audible-range communication at a distance, and exteroceptive sensing. We demonstrate a subset of the decentralized collective behaviors that these functions make possible with multi-robot hardware implementations. The use of acoustic waves in this domain is shown to provide distinct advantages over existing solutions.

Animals are capable of extreme agility, yet understanding their complex dynamics, which have ecological, biomechanical and evolutionary implications, remains challenging. Being able to study this incredible agility will be critical for the development of next-generation autonomous legged robots. In particular, the cheetah (acinonyx jubatus) is supremely fast and maneuverable, yet quantifying its wholebody 3D kinematic data during locomotion in the wild remains a challenge, even with new deep learning-based methods. In this work we present an extensive dataset of free-running cheetahs in the wild, called AcinoSet, that contains 119, 490 frames of multi-view synchronized high-speed video footage, camera calibration files and 7, 588 human-annotated frames. We utilize markerless animal pose estimation to provide 2D keypoints. Then, we use three methods that serve as strong baselines for 3D pose estimation tool development: traditional sparse bundle adjustment, an Extended Kalman Filter, and a trajectory optimization-based method we call Full Trajectory Estimation. The resulting 3D trajectories, human-checked 3D ground truth, and an interactive tool to inspect the data is also provided. We believe this dataset will be useful for a diverse range of fields such as ecology, neuroscience, robotics, biomechanics as well as computer vision. Code and data can be found at: https://github.com/African-Robotics-Unit/AcinoSet.

One simplifying assumption made in the existing and well-performing multi-robot systems is that the robots are single-tasking: each robot operates on a single task at any time. While this assumption is innocent to make in situations with sufficient resources such that robots can work independently, it becomes a restriction when they must share capabilities. In this paper, we consider multitasking robots with multi-robot tasks. Given a set of tasks, each achievable by a coalition of robots, our approach allows the coalitions to overlap by exploiting task synergies based on the physical constraints required to maintain these coalitions. The key contribution is a general and flexible framework that extends the current multi-robot systems to enable multitasking. The proposed approach is inspired by the information invariant theory, which orients around the equivalence of different information requirements. We map physical constraints to information requirements in our work, thereby allowing task synergies to be identified by reasoning about the relationships between such requirements. We show that our algorithm is sound and complete. Simulation results show its effectiveness under resource-constrained situations and in handling challenging scenarios in a realistic UAV simulator.

In the industrial domain, the application of any system as a safety function has to follow strict rules and requirements, defined in safety standards such as ISO 13849 [1] and ISO 13855 [2]. Two core requirements are an extremely low rate of dangerous errors and an upper limit for the response time of the system (hard real-time requirement). Current approaches in the field of human pose estimation achieve neither of both.In our work we approach the second requirement by introducing a general procedure to achieve hard real-time capability for interchangeable 3D human pose estimation systems. We use the detections of the pose estimation to model the human as 3D volume consisting of spheres and spherical cones. To bridge the time between arriving detections, the volume is adjusted in a way that ensures coherence, continuity and, most important, conformance with necessary surcharges from safety standard ISO 13855 [2]. Low and fixed computational cost for the adaption makes the procedure real-time capable. Our modelling approach using spheres and spherical cones also allows distance calculations at low and fixed computational costs, e.g. between human and robot. We show the benefit of our approach via human-robot distance calculation experiments, outperforming a safety-certified laser scanner for most of the time.

Vision-based localization and mapping in the agricultural environment is challenging due to the unstructured scene with unstable features, illumination variations, bumpy roads, and dynamic environmental objects. To address these challenges, we propose an accurate and robust stereo direct visual odometry system with modifications on Stereo-DSO. We firstly select some well-matched static stereo points in the latest keyframe to improve the accuracy of inverse depth calculation for tracking. The inverse depth can further distinguish close objects from background, which will avoid large and far-away scene objects in keyframe determination. To boost efficiency and accuracy at the tracking stage, we propose a point selection method to sample map points and remove outliers. Furthermore, altitude smoothness verification with a local flat ground assumption and recovery method for tracking failure on bumpy roads are proposed to improve the system’s robustness. Finally, a far-away keyframe is reserved in the sliding window to alleviate the orientation drift since the agricultural robots usually move straightly following the crop row. Our system achieved new state-of-the-art results on Flourish dataset and the recently released Rosario dataset.

Scale ambiguity is a fundamental problem in monocular visual odometry. Typical solutions include loop closure detection and environment information mining. For applications like self-driving cars, loop closure is not always available, hence mining prior knowledge from the environment becomes a more promising approach. In this paper, with the assumption of a constant height of the camera above the ground, we develop a light-weight scale recovery framework leveraging an accurate and robust estimation of the ground plane. The framework includes a ground point extraction algorithm for selecting high-quality points on the ground plane, and a ground point aggregation algorithm for joining the extracted ground points in a local sliding window. Based on the aggregated data, the scale is finally recovered by solving a least-squares problem using a RANSAC-based optimizer. Sufficient data and robust optimizer enable a highly accurate scale recovery. Experiments on the KITTI dataset show that the proposed framework can achieve state-of-the-art accuracy in terms of translation errors, while maintaining competitive performance on the rotation error. Due to the light-weight design, our framework also demonstrates a high frequency of 20 Hz on the dataset.

Reduction of combinatorial filters involves compressing state representations that robots use. Such optimization arises in automating the construction of minimalist robots. But exact combinatorial filter reduction is an NP-complete problem and all current techniques are either inexact or formalized with exponentially many constraints. This paper proposes a new formalization needing only a polynomial number of constraints, and characterizes these constraints in three different forms: nonlinear, linear, and conjunctive normal form. Empirical results show that constraints in conjunctive normal form capture the problem most effectively, leading to a method that outperforms the others. Further examination indicates that a substantial proportion of constraints remain inactive during iterative filter reduction. To leverage this observation, we introduce just-in-time generation of such constraints, which yields improvements in efficiency and has the potential to minimize large filters.

Probabilistic volumetric mapping (PVM) represents a 3D environmental map for an autonomous robotic navigational task. A popular implementation such as Octomap is widely used in the robotics community for such a purpose. The Octomap relies on an octree to represent a PVM and its main bottleneck lies in massive ray-shooting to determine the occupancy of the underlying volumetric voxel grids.In this paper, we propose GPU-based ray shooting to drastically improve the ray shooting performance in Octomap. Our main idea is based on the use of recent ray-tracing RTX GPU, mainly designed for real-time photo-realistic computer graphics and the accompanying graphics API, known as DXR. Our ray-shooting first maps leaf-level voxels in the given octree to a set of axis-aligned bounding boxes (AABBs) and employ massively parallel ray shooting on them using GPUs to find free and occupied voxels. These are fed back into the CPU to update the voxel occupancy and restructure the octree. In our experiments, we have observed more than three-orders-of-magnitude performance improvement in terms of ray shooting using ray-tracing RTX GPU over a state-of-the-art Octomap CPU implementation, where the benchmarking environments consist of more than 77K points and 25K~34K voxel grids.

Factor graph chains– the special case of a factor graph in which there are no potentials connecting non-adjacent nodes– arise naturally in many robotics problems. Importantly, they are often part of an inner loop in trajectory optimization and estimation problems, and so applications can be very sensitive to the performance of a solver.Of course, it is well-known that factor graph chains have an O(N) solution, but an actual solution is often left as "an exercise to the reader"… with the inevitable consequence that few (if any) efficient solutions are readily available.In this paper, we carefully derive the solution while keeping track of the specific block structure that arises, we work through a number of practical implementation challenges, and we highlight additional optimizations that are not at first apparent. An easy-to-use and self-contained solver is provided in C, which outperforms the AprilSAM general-purpose sparse matrix factorization library by a factor of 7.3x even without specialized block operations.The name AXLE reflects the names of the key matrices involved (the approach here solves the linear problem AX = E by factoring A as LLT), while also reflecting its key application in kino-dynamic trajectory estimation of vehicles with axles.

Pedestrian trajectory prediction is a critical yet challenging task especially for crowded scenes. We suggest that introducing an attention mechanism to infer the importance of different neighbors is critical for accurate trajectory prediction in scenes with varying crowd size. In this work, we propose a novel method, AVGCN, for trajectory prediction utilizing graph convolutional networks (GCN) based on human attention (A denotes attention, V denotes visual field constraints). First, we train an attention network that estimates the importance of neighboring pedestrians, using gaze data collected as subjects perform a bird’s eye view crowd navigation task. Then, we incorporate the learned attention weights modulated by constraints on the pedestrian’s visual field into a trajectory prediction network that uses a GCN to aggregate information from neighbors efficiently. AVGCN also considers the stochastic nature of pedestrian trajectories by taking advantage of variational trajectory prediction. Our approach achieves state-of-the-art performance on several trajectory prediction benchmarks, and the lowest average prediction error over all considered benchmarks.

Recognizing human emotion/expressions automatically is quite an expected ability for intelligent robotics, as it can promote better communication and cooperation with humans. Current deep-learning-based algorithms may achieve impressive performance in some lab-controlled environments, but they always fail to recognize the expressions accurately for the uncontrolled in-the-wild situation. Fortunately, facial action units (AU) describe subtle facial behaviors, and they can help distinguish uncertain and ambiguous expressions. In this work, we explore the correlations among the action units and facial expressions, and devise an AU-Expression Knowledge Constrained Representation Learning (AUE-CRL) framework to learn the AU representations without AU annotations and adaptively use representations to facilitate facial expression recognition. Specifically, it leverages AU-expression correlations to guide the learning of the AU classifiers, and thus it can obtain AU representations without incurring any AU annotations. Then, it introduces a knowledge-guided attention mechanism that mines useful AU representations under the constraint of AU-expression correlations. In this way, the framework can capture local discriminative and complementary features to enhance facial representation for facial expression recognition. We conduct experiments on the challenging uncontrolled datasets to demonstrate the superiority of the proposed framework over current state-of-the-art methods. Codes and trained models are available at https://github.com/HCPLab-SYSU/AUE-CRL.

The energy of ocean waves is the key distinguishing factor of marine environments compared to other aquatic environments such as lakes and rivers. Waves significantly affect the dynamics of marine vehicles; hence it is imperative to consider the dynamics of vehicles in waves when developing efficient control strategies for autonomous surface vehicles (ASVs). However, most marine simulators available open-source either exclude dynamics of vehicles in waves or use methods with high computational overhead. This paper presents ASVLite, a computationally efficient ASV simulator that uses frequency domain analysis for wave force computation. ASVLite is suitable for applications requiring low computational overhead and high run-time performance. Our tests on a Raspberry Pi 2 and a mid-range desktop computer show that the simulator has a high run-time performance to efficiently simulate irregular waves with a component wave count of up to 260 and large-scale swarms of up to 500 ASVs.

Human-robot collaboration frequently requires extensive communication, e.g., using natural language and gesture. Augmented reality (AR) has provided an alternative way of bridging the communication gap between robots and people. However, most current AR-based human-robot communication methods are unidirectional, focusing on how the human adapts to robot behaviors, and are limited to single-robot domains. In this paper, we develop AR for Robots Collaborating with a Human (ARROCH), a novel algorithm and system that supports bidirectional, multi-turn, human-multi-robot communication in indoor multi-room environments. The human can see through obstacles to observe the robots’ current states and intentions, and provide feedback, while the robots’ behaviors are then adjusted toward human-multi-robot teamwork. Experiments have been conducted with real robots and human participants using collaborative delivery tasks. Results show that ARROCH outperformed a standard non-AR approach in both user experience and teamwork efficiency. In addition, we have developed a novel simulation environment using Unity (for AR and human simulation) and Gazebo (for robot simulation). Results in simulation demonstrate ARROCH’s superiority over AR-based baselines in human-robot collaboration.

Learning from Demonstration (LfD) enables novice users to teach robots new skills. However, many LfD methods do not facilitate skill maintenance and adaptation. Changes in task requirements or in the environment often reveal the lack of resiliency and adaptability in the skill model. To overcome these limitations, we introduce ARC-LfD: an Augmented Reality (AR) interface for constrained Learning from Demonstration that allows users to maintain, update, and adapt learned skills. This is accomplished through in-situ visualizations of learned skills and constraint-based editing of existing skills without requiring further demonstration. We describe the existing algorithmic basis for this system as well as our Augmented Reality interface and the novel capabilities it provides. Finally, we provide three case studies that demonstrate how ARC-LfD enables users to adapt to changes in the environment or task which require a skill to be altered after initial teaching has taken place.

Classical navigation systems typically operate using a fixed set of hand-picked parameters (e.g. maximum speed, sampling rate, inflation radius, etc.) and require heavy expert re-tuning in order to work in new environments. To mitigate this requirement, it has been proposed to learn parameters for different contexts in a new environment using human demonstrations collected via teleoperation. However, learning from human demonstration limits deployment to the training environment, and limits overall performance to that of a potentially-suboptimal demonstrator. In this paper, we introduce APPLR, Adaptive Planner Parameter Learning from Reinforcement, which allows existing navigation systems to adapt to new scenarios by using a parameter selection scheme discovered via reinforcement learning (RL) in a wide variety of simulation environments. We evaluate APPLR on a robot in both simulated and physical experiments, and show that it can outperform both a fixed set of hand-tuned parameters and also a dynamic parameter tuning scheme learned from human demonstration.