ICRA 2022 论文列表

In this paper, we examine the problem of rearranging many objects on a tabletop in a cluttered setting using overhand grasps. Efficient solutions for the problem, which capture a common task that we solve on a daily basis, are essential in enabling truly intelligent robotic manipulation. In a given instance, objects may need to be placed at temporary positions (“buffers”) to complete the rearrangement, but allocating these buffer locations can be highly challenging in a cluttered environment. To tackle the challenge, a two-step baseline planner is first developed, which generates a primitive plan based on inherent combinatorial constraints induced by start and goal poses of the objects and then selects buffer locations assisted by the primitive plan. We then employ the “lazy” planner in a tree search framework which is further sped up by adapting a novel preprocessing routine. Simulation experiments show our methods can quickly generate high-quality solutions and are more robust in solving large-scale instances than existing state-of-the-art approaches. source: github.com/arc-l/TRLB

Tracking of tissue in the surgical environment is often done via locating frame-to-frame keypoint correspondences, and then using these correspondences to warp a prior underlying model such as a spline, mesh, or embedded deformation. We introduce a novel learned model which takes keypoint correspondences as input and enables a prior-free estimation of deformation at any location. For fast point tracking, our model allows for sparse queries, unlike dense grid based CNNs, which run on full images. Our model begins with a novel graph-based point refinement scheme which refines matched keypoints, updating their features and movement instead of discarding possible outliers. Then, we use these refined matches to learn a novel neural implicit representation for estimating movement of any location given its k-nearest neighbor (k-NN) keypoints. We name our implicit deformation model KINFlow (k-NN implicit neural flow). We demonstrate the performance of KINFlow photometrically on three different datasets. KINFlow is the first model to use a graph network to estimate flow of arbitrary query points, and can estimate movement of 1024 points in under 3 ms.

One of the fundamental challenges in realizing the potential of legged robots is generating plans to traverse challenging terrains. Control actions must be carefully selected so the robot will not crash or slip. The high dimensionality of the joint space makes directly planning low-level actions from onboard perception difficult, and control stacks that do not consider the low-level mechanisms of the robot in planning are ill-suited to handle fine-grained obstacles. One method for dealing with this is selecting footstep locations based on terrain characteristics. However, incorporating robot dynamics into footstep planning requires significant computation, much more than in the quasi-static case. In this work, we present an LSTM-based planning framework that learns probability distributions over likely footstep locations using both terrain lookahead and the robot's dynamics, and leverages the LSTM's sequential nature to find footsteps in linear time. Our framework can also be used as a module to speed up sampling-based planners. We validate our approach on a simulated one-legged hopper over a variety of uneven terrains.

We present a fast collision checking/avoidance algorithm for collaborative robot arms that work in close proximity. We formulate forward kinematics and separating distance function using DH convention and Taylor models (the tight enclosure of a function), and then compute their tight bounds for determining interference between robot arms. Our algorithm allows the collaborative robot arms to perform fine and dexterous tasks in close spatial proximity.

Future robotic deployments will require robots to be able to repeatedly solve a variety of tasks in application domains. Task and motion planning addresses complex robotic problems that combine discrete reasoning over states and actions and geometric interactions during action executions. Moving beyond deterministic settings, stochastic actions can be handled by modeling the problem as a Markov Decision Process. The underlying probabilities however are typically hard to model since failures might be caused by hardware imperfections, sensing noise, or physical interactions. We pro-pose a framework to address a task and motion planning setting where actions can fail during execution. To achieve a task goal actions need to be computed and executed despite failures. The robot has to infer which actions are robust and for each new problem effectively choose a solution that reduces expected execution failures. The key idea is to continually recover and refine the underlying beliefs associated with actions across multiple different problems in the domain. Our proposed method can find solutions that reduce the expected number of discrete, executed actions. Results in physics-based simulation indicate that our method outperforms baseline replanning strategies to deal with failing executions.

In this paper, we fabricated a soft sensor based on PEDOT:PSS for thin film structure. The developed soft sensor can measure the contraction force at real time to be embedded in a modular bio-actuator [1]. The modular actuator generated contraction forces at 0.3 mN when applying electric pulse stimulation. To measure millinewton contraction forces and make a built in sensor, we fabricated a soft sensor using PEDOT:PSS-PDMS film. To verify that the sensor can measure the force of the actuator and can be integrated to the actuator, we analyzed characteristic of the sensor. First, we measure Young's modulus of the sensor and compare them with the bio-actuator. From the previous research [2], the Young's modulus of the bio-actuator and sensor were 45.8 kPa and 165 kPa, respectively. In addition, we simulated the sensors to estimate the change of the displacement according to the applied force. Next, we have experiments by stretching sensors using stepping motor to measure the resistance change of the sensor. From the simulation data, the displacement change is 23 µm when applying 0.3 mN of forces and then we detect the displacement change smaller than is 20 µm from the experiments. Finally, we analyzed the movement of the bio-actuator when applying stimulation using high speed camera and time response of the developed sensor. The actuator was contracted to the maximum after 150 ms from the electrical stimulation and the sensor detected the repeated motion at 10 Hz without time delay. As a result, the proposed sensor can measure the force of bioactuator at real time.

We present a novel framework for floor plan-based, full six degree-of-freedom LiDAR localization. Our approach relies on robust ceiling and ground plane detection, which solves part of the pose and supports the segmentation of vertical structure elements such as walls and pillars. Our core contribution is a novel nearest neighbour data structure for an efficient look-up of nearest vertical structure elements from the floor plan. The registration is realized as a pair-wise regularized windowed pose graph optimization. Highly efficient, accurate and drift-free long-term localization is demonstrated on multiple scenes.

Grasping unknown objects with multi-fingered hands at high success rates and in real-time is an unsolved problem. Existing methods are limited in the speed of grasp synthesis or the ability to synthesize a variety of grasps from the same observation. We introduce Five-finger Hand Net (FFHNet), an ML model which can generate a wide variety of high-quality multi-fingered grasps for unseen objects from a single view. Generating and evaluating grasps with FFHNet takes only 30ms on a commodity GPU. To the best of our knowledge, FFHNet is the first ML-based real-time system for multi-fingered grasping with the ability to perform grasp inference at 30 frames per second (FPS). For training, we synthetically generate 180k grasp samples for 129 objects. We are able to achieve 91% grasping success for unknown objects in simulation and we demonstrate the model's capabilities of synthesizing high-quality grasps also for real unseen objects.

In this paper, we propose a novel consistent state estimator design for visual-inertial systems. Motivated by first-estimates Jacobian (FEJ) based estimators - which uses the first-ever estimates as linearization points to preserve proper observability properties of the linearized estimator thereby improving the consistency - we carefully model measurement linearization errors due to its Jacobian evaluation and propose a methodology which still leverages FEJ to ensure the estimator's observability properties, but additionally explicitly compensate for linearization errors caused by poor first estimates. We term this estimator FEJ2, which directly addresses the discrepancy between the best Jacobian evaluated at the latest state estimate and the first-estimates Jacobian evaluated at the first-time-ever state estimate. We show that this process explicitly models that the FEJ used is imperfect and thus contributes additional error which, as in FEJ2, should be modeled and consistently increase the state covariance during update. The proposed FEJ2 is evaluated against state-of-the-art visual-inertial estimators in both Monte-Carlo simulations and real-world experiments, which has been shown to outperform existing methods and to robustly handle poor first estimates and high measurement noises.

It is well known that visual SLAM systems based on dense matching are locally accurate but are also susceptible to long-term drift and map corruption. In contrast, feature matching methods can achieve greater long-term consistency but can suffer from inaccurate local pose estimation when feature information is sparse. Based on these observations, we propose an RGB-D SLAM system that leverages the advantages of both approaches: using dense frame-to-model odometry to build accurate sub-maps and on-the-fly feature-based matching across sub-maps for global map optimisation. In addition, we incorporate a learning-based loop closure component based on 3-D features which further stabilises map building. We have evaluated the approach on indoor sequences from public datasets, and the results show that it performs on par or better than state-of-the-art systems in terms of map reconstruction quality and pose estimation. The approach can also scale to large scenes where other systems often fail.

We demonstrate a multi-lidar calibration frame-work for large mobile platforms that jointly calibrate the extrinsic parameters of non-overlapping Field-of-View (FoV) lidar sensors, without the need for any external calibration aid. The method starts by estimating the pose of each lidar in its corresponding sensor frame in between subsequent timestamps. Since the pose estimates from the lidars are not necessarily synchronous, we first align the poses using a Dual Quaternion (DQ) based Screw Linear Interpolation. Afterward, a Hand-Eye based calibration problem is solved using the DQ-based formulation to recover the extrinsics. Furthermore, we verify the extrinsics by matching chosen lidar semantic features, obtained by projecting the lidar data into the camera perspective after time alignment using vehicle kinematics. Experimental results on the data collected from a Scania vehicle [~ 1 Km sequence] demonstrate the ability of our approach to obtain better calibration parameters than the provided vehicle CAD model calibration parameters. This setup can also be scaled to any combination of multiple lidars.

Autonomous exploration and mapping of unknown terrains employing single or multiple robots is an essential task in mobile robotics and has therefore been widely investigated. Nevertheless, given the lack of unified data sets, metrics, and platforms to evaluate the exploration approaches, we develop an autonomous robot exploration benchmark en-titled Explore-Bench. The benchmark involves various explo-ration scenarios and presents two types of quantitative metrics to evaluate exploration efficiency and multi-robot cooperation. Explore-Bench is extremely useful as, recently, deep rein-forcement learning (DRL) has been widely used for robot exploration tasks and achieved promising results. However, training DRL-based approaches requires large data sets, and additionally, current benchmarks rely on realistic simulators with a slow simulation speed, which is not appropriate for training exploration strategies. Hence, to support efficient DRL training and comprehensive evaluation, the suggested Explore-Bench designs a 3-level platform with a unified data flow and 12 × speed-up that includes a grid-based simulator for fast evaluation and efficient training, a realistic Gazebo simulator, and a remotely accessible robot testbed for high-accuracy tests in physical environments. The practicality of the proposed benchmark is highlighted with the application of one DRL-based and three frontier-based exploration approaches. Fur-thermore, we analyze the performance differences and provide some insights about the selection and design of exploration methods. Our benchmark is available at https://github.com/efc-robot/Explore-Bench.

Despite recent progress of robotic exploration, most methods assume that drift-free localization is available, which is problematic in reality and causes severe distortion of the reconstructed map. In this work, we present a systematic exploration mapping and planning framework that deals with drifted localization, allowing efficient and globally consistent reconstruction. A real-time re-integration-based mapping approach along with a frame pruning mechanism is proposed, which rectifies map distortion effectively when drifted localization is corrected upon detecting loop-closure. Besides, an exploration planning method considering historical viewpoints is presented to enable active loop closing, which promotes a higher opportunity to correct localization errors and further improves the mapping quality. We evaluate both the mapping and planning methods as well as the entire system comprehensively in simulation and real-world experiments, showing their effectiveness in practice. The implementation of the proposed method will be made open-source for the benefit of the robotics community.

Self-driving cars must detect other traffic participants like vehicles and pedestrians in 3D in order to plan safe routes and avoid collisions. State-of-the-art 3D object detectors, based on deep learning, have shown promising accuracy but are prone to over-fit domain idiosyncrasies, making them fail in new environments-a serious problem for the robustness of self-driving cars. In this paper, we propose a novel learning approach that reduces this gap by fine-tuning the detector on high-quality pseudo-labels in the target domain - pseudo-labels that are automatically generated after driving based on replays of previously recorded driving sequences. In these replays, object tracks are smoothed forward and backward in time, and detections are interpolated and extrapolated-crucially, leveraging future information to catch hard cases such as missed detections due to occlusions or far ranges. We show, across five autonomous driving datasets, that fine-tuning the object detector on these pseudo-labels substantially reduces the domain gap to new driving environments, yielding strong improvements detection reliability and accuracy.

Reinforcement Learning is rapidly establishing itself as the foremost choice for optimization of sequential autonomous decision-making problems. Encumbered by its sample inefficiency, the extension of the field to large state space and dynamic environments remains an open problem. We present a novel concept that exploits abstract spatial symmetry in complex environments for extending the skills of naïvely trained agents in local abstractions of the environment. The concept of EASE (Exploitation of Abstract Symmetry of Environments), when incorporated, improves the sample efficiency of traditional reinforcement learning algorithms. The presented work exemplifies the concept of EASE by presenting three distinct settings; EASE with heuristics-based planning, EASE with learning from demonstrations and EASE with state-space abstraction and proposes a novel algorithm for each setting.

Fast, autonomous flight in unstructured, cluttered environments such as forests is challenging because it requires the robot to compute new plans in realtime on a computationally-constrained platform. In this paper, we enable this capability with a search-based planning framework that adapts sampling density in realtime to find dynamically-feasible plans while remaining computationally tractable. A paramount challenge in search-based planning is that dense obstacles both necessitate large graphs (to guarantee completeness) and reduce the efficiency of graph search (as heuristics become less accurate). To address this, we develop a planning framework with two parts: one that maximizes planner completeness for a given graph size, and a second that dynamically maximizes graph size subject to computational constraints. This framework is enabled by motion planning graphs that are defined by a single parameter-dispersion-which quantifies the maximum trajectory cost to reach an arbitrary state from the graph. We show through real and simulated experiments how the dispersion can be adapted to different environments in realtime, allowing operation in environments with varying density. The simulated experiment demonstrates improved performance over a baseline search-based planning algorithm. We also demonstrate flight speeds of up to 2.5m/s in real-world cluttered pine forests.

This paper reports on preliminary experimental results of recently proposed mechanism kinematics for a legged robot. The proposed kinematics creates a mapping from a series-elastic actuator to a foot motion that includes a pair of singularities within a fully rotatable kinematic circuit. Such a circuit is less common and only possible with certain multi-loop linkages. A slice of the configuration space displaying series-elastic rotation versus linear foot motion presents a characteristic “S” shape, motivating the name S-curve kinematics. Our experimental results show that S-curve kinematics can enhance the energetic output of a series-elastic actuator in a hopping task versus the usage of a conventional rotary-to-linear mechanism. This is possible because S-curve kinematics enable elastic energy storage outside of stance that is released through a mechanical reflex. Compared to a conventional rotary-to-linear actuator, S-curve kinematics demonstrated up to a 4x increase in kinetic output.

Handed Shearing Auxetics (HSA) are a promising structure for making electrically driven robots with distributed compliance that convert a motors rotation and torque into extension and force. These structures expand and contract by changing an internal angle between links, the evolution of the structure as this angle changes is known as the auxetic trajectory. We overcome past limitations on the range of actuation, blocked force, and stiffness by focusing on two key design parameters: the point of an HSA's auxetic trajectory that is energetically preferred, and the number of cells along the HSAs length. Modeling the HSA as a programmable spring, we characterize the effect of both on blocked force, minimum energy length, spring constant, angle range and holding torque. We also examined the effect viscoelasticity has on actuation forces over time. By varying the preferred auxetic trajectory point, we were able to make actuators that can push, pull, or do both. We expanded the range of forces possible from 5 N to 150 N, and the range of stiffness from 2 N/mm to 89 N/mm. For a fixed point on the auxetic trajectory, we found decreasing length can improve force output, at the expense of needing higher torques, and having a shorter throw. We also found that the viscoelastic effects can limit the amount of force a 3D printed HSA can apply over time.

A predictive model for mobility systems capable of understanding the trajectory a user intends to follow in the environment is proposed. Understanding user intention is paramount for any shared-control navigation strategy between a user and an active robotic agent. Equally important however is being able to go beyond simple sample generation to assign probabilistic meaning to the set of possible future trajectories, so most likely scenarios can be assumed. The framework estimates a distribution over possible intentions, proposing a novel generative model predicated on Normalizing Flows which accounts for past behaviours, as traditionally reported in the literature, but also incorporates visual scene information. As the model permits trajectories to be assigned exact likelihoods, tractable density estimates can be readily exploited to finalize an executable intention. Baseline comparisons with the publicly available and widely used KITTI navigational dataset show significant improvements (up to 11.08%) with respect to traditional metrics such as Average and Final Displacement Errors. A novel metric that stands independent of the number of samples is also proposed as a more fitting comparison for future works.