IROS 2022 Papers

Place recognition is a fundamental component of robotics, and has seen tremendous improvements through the use of deep learning models in recent years. Networks can experience significant drops in performance when deployed in unseen or highly dynamic environments, and require additional training on the collected data. However naively fine-tuning on new training distributions can cause severe degradation of performance on previously visited domains, a phenomenon known as catastrophic forgetting. In this paper we address the problem of incremental learning for point cloud place recognition and introduce InCloud, a structure-aware distillation-based approach which preserves the higher-order structure of the network's embedding space. We introduce several challenging new benchmarks on four popular and large-scale LiDAR datasets (Oxford, MulRan, In-house and KITTI) showing broad improvements in point cloud place recognition performance over a variety of network architectures. To the best of our knowledge, this work is the first to effectively apply incremental learning for point cloud place recognition. Data pre-processing, training and evaluation code for this paper can be found at https://github.com/csiro-robotics/InCloud.

In this work, we investigate the problem of incre-mentally solving constrained non-linear optimization problems formulated as factor graphs. Prior incremental solvers were either restricted to the unconstrained case or required periodic batch relinearizations of the objective and constraints which are expensive and detract from the online nature of the algorithm. We present InCOpt, an Augmented Lagrangian-based incremental constrained optimizer that views matrix operations as message passing over the Bayes tree. We first show how the linear system, resulting from linearizing the constrained objective, can be represented as a Bayes tree. We then propose an algorithm that views forward and back substitutions, which naturally arise from solving the Lagrangian, as upward and downward passes on the tree. Using this formulation, In-COpt can exploit properties such as fluid/online relinearization leading to increased accuracy without a sacrifice in runtime. We evaluate our solver on different applications (navigation and manipulation) and provide an extensive evaluation against existing constrained and unconstrained solvers.

This paper presents a novel method of realizing in-hand manipulation inspired by the peristaltic motion of a large-sized caterpillar. The sharp contrast between the proposed soft-bodied finger and the conventional hard/rigid robotic ones is peristaltic motion with compression and bending deformations. The design is based on the biological fact that large-size caterpillars (e.g., Bombyx mori) utilize bending and compression/extension of the body to produce crawling locomotion. Exploiting the multi-modal deformations, we demonstrated that the prototype hand comprising two proposed fingers could rotate and transport grasped objects. We also observed that the time gap of two-finger motion is required to stabilize in-hand manipulation. The design can provide new insights into designing a gripper inspired by soft-bodied creatures.

While sampling-based approaches have made significant progress, motion planning with dynamics still poses significant challenges as the planner has to generate not only collision-free but also dynamically-feasible trajectories that enable the robot to reach its goal. To improve the efficiency of sampling-based motion planners, this paper develops a framework, termed Motion-Planning Runtime Prediction (MPRP), that relies on machine learning to train models to predict the expected runtime of a planner. When solving a new motion-planning problem, the trained model is then incorporated into the motion planner to more effectively guide the search toward parts of the state space that are associated with low expected runtime predictions. This paper applies the MPRP framework to state-of-the-art sampling-based motion planners to obtain new planners, which are shown to be significantly faster.

As most existing works of single-view 3D reconstruction aim at learning the better mapping functions to directly transform the 2D observation into the corresponding 3D shape for achieving state-of-the-art performance, there often comes a potential concern on having the implicit bias towards the seen classes learnt in their models (i.e. reconstruction intertwined with the classification) thus leading to poor generalizability for the unseen object categories. Moreover, such implicit bias typically stemmed from adopting the object-centered coordinate in their model designs, in which the reconstructed 3D shapes of the same class are all aligned to the same canonical pose regardless of different view-angles in the 2D observations. To this end, we propose an end-to-end framework to reconstruct the 3D mesh from a single image, where the reconstructed mesh is not only view-centered (i.e. its 3D pose respects the viewpoint of the 2D observation) but also preliminarily represented as a composition of volumetric 3D primitives before being further deformed into the fine-grained mesh to capture the shape details. In particular, the usage of volumetric primitives is motivated from the assumption that there generally exists some similar shape parts shared across various object categories, learning to estimate the primitive-based 3D model thus becomes more generalizable to the unseen categories. Furthermore, we advance to propose a novel mesh augmentation strategy, CvxRearrangement, to enrich the distribution of training shapes, which contributes to increasing the robustness of our proposed model and achieves better generalization. Extensive experiments demonstrate that our proposed method provides superior performance on both unseen and seen classes in comparison to several representative baselines of single-view 3D reconstruction.

Radar odometry may provide valuable input for surface vessels in several marine applications. The vulnerability of global positioning satellite systems to jamming and spoofing motivates the search for alternatives. In this work, we investigate the feasibility of W-band frequency modulated continuous wave radars in marine settings for odometry. A method to model radar resolution is presented and is further extend to include multiple radar frames. Numerical implementation relies on concepts from Lie theory. The proposed methods were evaluated on datasets collected from a ferry in a harbour area, where the average relative translation error was reduced to 3.07% compared to 14.2% of a baseline method.

Tracking the 3D motion of agile animals in the wild will enable new insight into the design of robotic controllers. However, in-field 3D pose estimation of high-speed wildlife such as cheetahs is still a challenge [1]. In this work, we aim to solve two of these challenges: unnatural pose estimates during highly occluded sequences and synchronization error between multi-view data. We expand on our previous Full Trajectory Estimation (FTE) method with two significant additions: Pairwise FTE (PW-FTE) and Shutter-delay FTE (SD-FTE). The PW-FTE expands on image-dependent pairwise terms, produced by a convolutional neural network (CNN), to infer occluded 2D keypoints, while SD-FTE uses shutter delay estimation to correct the synchronization error. Lastly, we combine both methods into PW-SD-FTE and perform a quantitative and qualitative analysis on a subset of AcinoSet, the video dataset of rapid and agile motions of cheetahs. We found that SD-FTE has significant benefits in tracking the position of the cheetah in the world frame, while PW-FTE provided a more robust 3D pose estimate during events of high occlusion. The PW-SD-FTE was found to retain both advantages, resulting in an improved baseline for AcinoSet. Code and data can be found at https://github.com/African-Robotics-Unit/AcinoSet/tree/pw_sd_fte.

This paper studies the effects of replacing pin-joint knees in passive dynamic bipedal walkers with biomimetic four-bar knees. The kinetic model of the four-bar knees is presented in detail, and an analytical model of the passive walking dynamics is derived. The resulting four-bar kneed biped is compared with a pin-joint kneed walker, for their passive walking performance. The geometry of the four-bar knees used in the study is based on human anatomical data. It is found that the biomimetic four-bar knee configuration works to the advantage of the biped, especially around the extended-knee singular position. The four-bar knees are found to overperform the pin-joint ones, resulting in significant reduction of peak impact loads and energetic expenditure.

Standing balance for legged robots can be achieved through regulating the center of pressure (ankle strategy), the angular momentum about the center of mass (hip strategy), and the magnitude of ground reaction force (variable height strategy). Prevalent reduced order models used to model legged robots at most only capture two of these strategies, and the contribution of the three available strategies is unclear. We propose a unified reduced order model that includes all three standing balance strategies and compared push recovery simulations of the unified model against existing balancing models using a nonlinear model predictive controller. We also developed a full body controller for a simple one legged balancing robot that tracked control from the reduced order models. For both the reduced order model and robot simulations, we found that the unified model could recover successfully from the largest pushes and yielded the smallest center of mass excursions. Between the hip and variable height strategies, the hip had the greatest effect on improving performance. Our results suggest that successful implementation of a unified reduced order model on physical robots would enable a simplified controller that takes advantage of available balancing strategies as needed to recover from larger push disturbances than feasible before.

In this work, an advanced quadruped robot with abundant kinematic loops and passive joints is introduced. Due to the existence of many closed chains, the robot dynamic model is quite complex, and is derived using the Gauss's principle of least constraint. To explicitly consider the loop-closure constraints, we propose a task-space inverse dynamics based approach to obtain the robot locomotion controller. Besides, to meet the demand of high frequency (≥ 500Hz) in controller, an alternative method is provided. It uses the projected dynamics to find an analytical mapping from the desired contact force to the desired torque of actuators under full consideration of passive joints and loop-closure constraints. The effectiveness and efficiency of the proposed algorithms in this paper have been validated by simulation with a reliable physical engine MuJoCo.

Meta Reinforcement Learning (Meta-RL) has seen substantial advancements recently. In particular, off-policy methods were developed to improve the data efficiency of Meta-RL techniques. Probabilistic embeddings for actor-critic \boldsymbol{RL}\boldsymbol{RL} (PEARL) is a leading approach for multi-MDP adaptation problems. A major drawback of many existing Meta-RL methods, including PEARL, is that they do not explicitly consider the safety of the prior policy when it is exposed to a new task for the first time. Safety is essential for many real world applications, including field robots and Autonomous Vehicles (AVs), In this paper, we develop the PEARL PLUS (PEARL+) algorithm, which optimizes the policy for both prior (pre-adaptation) safety and posterior (after-adaptation) performance. Building on top of PEARL, our proposed PEARL+ algorithm introduces a prior regularization term in the reward function and a new Q-network for recovering the state-action value under prior context assumptions, to improve the robustness to task distribution shift and safety of the trained network exposed to a new task for the first time. The performance of PEARL+ is validated by solving three safety-critical problems related to robots and AVs, including two MuJoCo benchmark problems. From the simulation experiments, we show that safety of the prior policy is significantly improved and more robust to task distribution shift compared to PEARL.

The difficulty associated with the coordinated locomotion of legged robots grows quickly as the number of joints increases. Although prior approaches have addressed this problem through sampling-based planners, learning-based techniques have recently been explored as a means to handle such complexity. Among these recent approaches are systems that utilize probabilistic graphical models in order to infer parameters for central pattern generators (CPGs) which enable the path-following locomotion of highly-articulated legged robots through unstructured terrain. This paper presents a novel formulation of a CPG parameter inference-based path-following controller. The new inference process and accompanying CPG formulation enforce oscillator convergence to the limit-cycle specified by the inferred parameters in addition to biasing towards parameters that quickly reach stable-state. This formulation is shown to improve the performance of CPG parameter inference-based path-following control for legged robots across a number of simulated and physical experiments.

This paper presents a motion planning strategy that utilizes the improved A -search guided tree to enable autonomous parking of a general 3-trailer with a car-like tractor. Different from the state-of-the-art state-lattice-based methods, where numerous motion primitives are necessary to ensure successful planning, our work allows quick off-lattice exploration to find a solution. Our treatment brings at least three advantages: fewer and lower design complexity of motion primitives, improved success rate, and increased path quality. Unlike on-lattice exploration, where the cost-to-go is obtained by querying a heuristic look-up table, off-lattice exploration entails the heuristic function being well-defined at off-lattice nodes. We train a neural network through reinforcement learning to model the maneuver costs of the trailer and use it as the heuristic value to better approximate the cost-to-go. Simulations demonstrate the effectiveness of the proposed method in terms of planning speed and path length.

Robot swarms have the potential to revolutionize areas ranging from warehouse management and agriculture to underwater and space exploration. However, there remains a substantial gap between theory and robot implementation. While algorithms might assume reliable communication, perfect sensing, and instantaneous cognition, most robots have lossy or even no communication, imperfect sensing, and limited cognition speed. In our previous work on implicit vision-based coordination, we demonstrated autonomous three-dimensional behaviors underwater by removing the need for radio communication between robots. Here we explore impressionist algorithms, capable of working with even more minimal information where traditional algorithms are prone to fail. Our case study focuses on classic flocking behaviors, where a robot swarm must coordinate group motion. We demonstrate that reliable alignment, dispersion, and milling can be achieved with only infrequent and imperfect sensory impressions. In simulation studies and theoretical analyses, we investigate the effect of systematically reducing spatial and temporal fidelity of individual information on the success metrics for the group; we also demonstrate physical experiments with Blueswarm robots using simple color detection. Our results show the potential of impressionist algorithms that operate on simpler neighborhood-awareness metrics and still achieve desired global goals.

Context information is important for instance segmentation on point clouds. Existing methods either only use local surroundings by stacking multiple convolution layers or use non-local methods to model long-range interactions. However, they usually directly operate on points which is an unstructured and low-level representation and is highly dependent on context. To address this issue, we propose an effective framework named Implicit-Part Context Aggregation (IPCA), which adopts implicit parts as an intermediate representation and achieves context aggregation through message passing along the implicit part graph. Specifically, we first organize unstructured points into geometrically consistent implicit parts and construct the implicit part graph according to the geometric adjacency. Then, an initial part embedding is extracted using the proposed Implicit Part Network (IPN) which can aggregate point features and capture the intrinsic geometric shape of the part. We further refine the part embedding by a graph reasoning module named Context Aggregation Network (CAN), which helps to make a more precise prediction by well exploiting the context information. Instance proposals are then generated by grouping implicit parts. Finally, we propose an additional step to attribute the entire instance proposal to a Semantic Criterion Net (SCN) to infer the semantics of the instance. The purpose is to correct the semantic prediction errors caused by not knowing the boundary and overall shape of the object in the previous steps. Extensive experiments on two large datasets, ScanNet and 3RScan, demonstrate the effectiveness of our method. To our knowledge, it yields the highest performance on the ScanNet test benchmark and its AP@50 is 9.5 points higher than the baseline. The code is available at https://github.com/xiaodongww/IPCA

Sound is one of the most informative and abundant modalities in the real world while being robust to sense without contacts by small and cheap sensors that can be placed on mobile devices. Although deep learning is capable of extracting information from multiple sensory inputs, there has been little use of sound for the control and learning of robotic actions. For unsupervised reinforcement learning, an agent is expected to actively collect experiences and jointly learn representations and policies in a self-supervised way. We build realistic robotic manipulation scenarios with physics-based sound simulation and propose the Intrinsic Sound Curiosity Module (ISCM). The ISCM provides feedback to a reinforcement learner to learn robust representations and to reward a more efficient exploration behavior. We perform experiments with sound enabled during pre-training and disabled during adaptation, and show that representations learned by ISCM outperform the ones by vision-only baselines and pre-trained policies can accelerate the learning process when applied to downstream tasks.

The recent momentum in aerial manipulation has led to an interest in developing virtual reality interfaces for aerial physical interaction tasks with simple, intuitive, and reliable control and perception. However, this requires the use of expensive subsystems and there is still a research gap between interface design, user evaluations and the effect on aerial manipulation tasks. Here, we present a methodology for low-cost available drone systems with a Unity-based interface for immersive FPV teleoperation. We applied our approach in a flight track where a cluttered environment is used to simulate a demanding aerial manipulation task inspired by forestry drones and canopy sampling. Through objective measures of teleoperation performance and subjective questionnaires, we found that operators performed worse using the FPV interface and had higher perceived levels of cognitive load when compared to traditional interface design. Additional analysis of physiological measures highlighted that objective stress levels and cognitive load were also influenced by task duration and perceived performance, providing an insight into what interfaces could target to support teleoperator requirements during aerial manipulation tasks.

Daily manipulation tasks are characterized by geometric primitives related to actions and object shapes. Such geometric descriptors are poorly represented by only using Cartesian coordinate systems. In this paper, we propose a learning approach to extract the optimal representation from a dictionary of coordinate systems to encode an observed movement/behavior. This is achieved by using an extension of Gaussian distributions on Riemannian manifolds, which is used to analyse a set of user demonstrations statistically, by considering multiple geometries as candidate representations of the task. We formulate the reproduction problem as a general optimal control problem based on an iterative linear quadratic regulator (iLQR), where the Gaussian distribution in the extracted coordinate systems are used to define the cost function. We apply our approach to object grasping and box opening tasks in simulation and on a 7-axis Franka Emika robot. The results show that the robot can exploit several geometries to execute the manipulation task and generalize it to new situations, by maintaining the invariant characteristics of the task in the coordinate system(s) of interest.

Automated excavation is promising to improve the safety and efficiency of excavators, and trajectory planning is one of the most important techniques. In this paper, we propose a two-stage method that integrates data-driven imitation learning and model-based trajectory optimization to generate optimal trajectories for autonomous excavators. We firstly train a deep neural network using demonstration data to mimic the operation patterns of human experts under various terrain states including their geometry shape and material type. Then, we use a stochastic trajectory optimization method to improve the trajectory generated by the neural network to guarantee kinematics feasibility, improve smoothness, satisfy hard constraints, and achieve desired excavation volumes. We test the proposed algorithm on a Franka robot arm equipped with a bucket end-effector. We further evaluate our method on different material types, such as sand and rigid blocks. The ex-perimental results show that the proposed two-stage algorithm by combining expert knowledge and model optimization can increase the excavation weights by up to 24.77% meanwhile with low variance.