ICRA 2021 Papers

We present an assistance system that reasons about a human’s intended actions during robot teleoperation in order to provide appropriate modifications on unintended behavior. Existing methods typically treat the human and control interface as a black box and assume the measured user input is noise-free, and use this signal to infer task-level human intent. We recognize that the signal measured through the interface is masked by the physical limitations of the user and the interface they are required to use. With this key insight, we model the human’s physical interaction with a control interface during robot teleoperation, and distinguish between interface-level intended and measured physical actions explicitly. By reasoning over the unobserved intentions using model-based inference techniques, our assistive system provides customized modifications on a user’s issued commands. We validate our algorithm both in simulation and with a 10-person human subject study in which we evaluate the performance of the proposed assistance paradigms. Our results show that the assistance paradigms helped to significantly reduce task completion time, number of mode switches, cognitive workload, and user frustration, and improve overall user satisfaction.

In the case of non-holonomic robot navigation, path planning algorithms such as Rapidly-exploring Random Tree (RRT) rarely provide feasible and smooth paths without the need of additional processing. Furthermore, in a transport context like power wheelchair navigation, passenger comfort should be a priority and influence path planning strategy. In this paper, we propose a local path planner which guarantees bounded curvature value and continuous Cubic Bézier piecewise curves connection. To simulate and test this Cubic Bézier local path planner, we developed a new RRT version (CBB-RRT*) which generates on-the fly comfortable path adapted to non-holonomic constraints.

The evaluation of robot capabilities to navigate human crowds is essential to conceive new robots intended to operate in public spaces. This paper initiates the development of a benchmark tool to evaluate such capabilities; our long term vision is to provide the community with a simulation tool that generates virtual crowded environment to test robots, to establish standard scenarios and metrics to evaluate navigation techniques in terms of safety and efficiency, and thus, to install new methods to benchmarking robots’ crowd navigation capabilities. This paper presents the architecture of the simulation tools, introduces first scenarios and evaluation metrics, as well as early results to demonstrate that our solution is relevant to be used as a benchmark tool.

A pair of Supernumerary Robotic Limbs (Super-Limbs) can brace the wearer’s upper body while they work at floor level, and support them during crawling. The SuperLimbs’ motion is synchronized with the operator to mimic natural human crawling. This synchronization relies on experimental data from the operator’s observed crawl. A method for predicting the phase difference between the SuperLimbs’ hand placement and the operators desired hand placement is developed and used to coordinate the SuperLimbs’ motion with the operator’s. The experimental data is also used to design the structure of the SuperLimbs to minimize their energy consumption during crawling and regulate their actuator’s temperatures.The SuperLimbs are designed to mimic the operator’s arms while they crawl, however factors such as their limited number of degrees of freedom (DoF), compared to natural limbs, and the compliance of the connection between the SuperLimbs and the operator means that the SuperLimbs’s may dynamically interact with the wearer differently from how their natural limbs do. A simple dynamic model is constructed to assess the energy consumed by the operator as they crawl with the SuperLimbs under different coordination patterns. An optimal coordination pattern is obtained from this assessment.

Deep learning algorithms are recognized as the most effective method for rectal tumor segmentation. However, since the multi-scale detailed feature information of rectal tumor cannot be fully extracted and applied, the segmentation and identification results of most algorithms are not always perfect. In this work, we introduce a Covariance Self-Attention Dual Path UNet (CSA-DPUNet), that is modified on the basis of UNet network and self-attention mechanism to improve network performance in feature processing and representation. The proposed network mainly makes two improvements. First, the UNet structure with single path is extended to dual paths (DPUNet). By broadening the network connections, our network is able to learn more local features with multiple contextual scales from CT images. Second, an improved criss-cross self-attention module is incorporated into DPUNet(CSA-DPUNet), instead of correlation method, we adopt covariance operation to calculate the attention weight map of self-attention mechanism, which can adaptively enhance feature combination and characterization ability. Experiments illustrate that our network called CSA-DPUNet can obviously improves the segmentation accuracy of rectal tumors, which brings 15.31%, 7.2%, 11.8%, and 9.5% improvement in Dice coefficient, P, R, F1, respectively compared with state-of-the-art. The above characteristics make the proposed CSA-DPUNet suitable for segmenting rectal tumor in practice.

This paper presents a real-time method for whole-body trajectory optimization of mobile manipulators in simplified dynamic and unstructured environments. Current trajectory optimization methods typically use decoupling of the mobile base and the robotic arm, which reduces flexibility in motion, does not scale to unstructured environments, and does not consider the future evolution of the environment, which is crucial to avoid dynamic obstacles. Given a goal configuration, such as waypoints generated by a global path planner, we formulate a receding horizon trajectory optimization minimizing the distance-to-target while avoiding collisions with static and dynamic obstacles. The presented method unifies the control of a robotic arm and a non-holonomic base to allow coupled trajectory planning. For collision avoidance, we propose to compute three convex regions englobing the robot's major body parts (i.e., base, shoulder-link and wrist-link) and thus reducing and limiting the number of inequality constraints, regardless of the number of obstacles in the environment. Moreover, our approach incorporates predicted trajectory information to smoothly, and in advance, avoid dynamic obstacles. The presented results show that trajectory optimization for the coupled system can reduce the total execution time by 48% and that applying the convex region generation for individual links allows keeping the computational costs low, even for complex scenarios, enabling onboard implementation.

This paper presents c2g-HOF networks which learn to generate cost-to-go functions for manipulator motion planning. The c2g-HOF architecture consists of a cost-to-go function over the configuration space represented as a neural network (c2g-network) as well as a Higher Order Function (HOF) network which outputs the weights of the c2g-network for a given input workspace. Both networks are trained end-to-end in a supervised fashion using costs computed from traditional motion planners. Once trained, c2g-HOF can generate a smooth and continuous cost-to-go function directly from workspace sensor inputs (represented as a point cloud in 3D or an image in 2D). At inference time, the weights of the c2g-network are computed very efficiently and near-optimal trajectories are generated by simply following the gradient of the cost-to-go function.We compare c2g-HOF with traditional planning algorithms for various robots and planning scenarios. The experimental results indicate that planning with c2g-HOF is significantly faster than other motion planning algorithms, resulting in orders of magnitude improvement when including collision checking. Furthermore, despite being trained from sparsely sampled trajectories in configuration space, c2g-HOF generalizes to generate smoother, and often lower cost, trajectories. We demonstrate cost-to-go based planning on a 7 DoF manipulator arm where motion planning in a complex workspace requires only 0.13 seconds for the entire trajectory.

This paper studies the problem of multi-robot cooperative visual-inertial localization where each robot is equipped with only a single camera and IMU. We develop two cooperative visual-inertial odometry (C-VIO) algorithms within the multi-state constraint Kalman filter (MSCKF) framework, in which each robot utilizes not only its own measurements but constraints of common features co-observed with its neighbors within the current sliding window in order to improve the localization accuracy. The first centralized-equivalent algorithm tracks the robot-to-robot cross correlations and prioritizes the pose accuracy while requiring full capacity communication among all the robots during update. The second distributed algorithm ignores the robot-to-robot cross correlations to obtain a scalable, robust and efficient fully distributed structure where each robot only keeps its own states and communicates with its neighbors, while a covariance intersection (CI)-based update strategy is leveraged to guarantee consistency. The proposed algorithms are validated extensively in both Monte-Carlo simulations and real-world datasets, and shown to be able to achieve better accuracy with competitive efficiency.

Driving energy consumption plays a major role in the navigation of mobile robots in challenging environments, especially if they are left to operate unattended under limited on-board power. This paper reports on first results of an energy-aware path planner, which can provide estimates of the driving energy consumption and energy recovery of a robot traversing complex uneven terrains. Energy is estimated over trajectories making use of a self-supervised learning approach, in which the robot autonomously learns how to correlate perceived terrain point clouds to energy consumption and recovery. A novel feature of the method is the use of 1D convolutional neural network to analyse the terrain sequentially in the same temporal order as it would be experienced by the robot when moving. The performance of the proposed approach is assessed in simulation over several digital terrain models collected from real natural scenarios, and is compared with a heuristic inclination-based energy model. We show evidence of the benefit of our method to increase the overall prediction r2 score by 66.8% and to reduce the driving energy consumption over planned paths by 5.5%.

This paper presents a new approach to Model Predictive Control for environments where essential, discrete variables are partially observed. Under this assumption, the belief state is a probability distribution over a finite number of states. We optimize a control-tree where each branch assumes a given state-hypothesis. The control-tree optimization uses the probabilistic belief state information. This leads to policies more optimized with respect to likely states than unlikely ones, while still guaranteeing robust constraint satisfaction at all times. We apply the method to both linear and non-linear MPC with constraints. The optimization of the control-tree is decomposed into optimization subproblems that are solved in parallel leading to good scalability for high number of state-hypotheses. We demonstrate the real-time feasibility of the algorithm on two examples and show the benefits compared to a classical MPC scheme optimizing w.r.t. one single hypothesis.

Control of an aerial manipulator is challenging due to the decentralized dynamics of the aerial vehicle and the robotic arm. It is generally complex to adjust the controller of the aerial manipulator when replacing a different robotic arm. This paper presents a flexible control scheme for a quadrotor-based aerial manipulator equipped with a replaceable robotic arm. To analyze the dynamic characteristics during grasping, the model of the aerial manipulator is decentralized including the models of a quadrotor and the centroid of an n-DOF robotic arm. The interaction effect of a moving robotic arm on the quadrotor is considered by analyzing the varying centroid of the robotic arm. Based on the modeling of the aerial manipulator, a control scheme integrating a linear model predictive control (LMPC) and a feedforward controller is presented to accurately control the motion of the aerial platform. The LMPC controls the aerial vehicle to follow the desired trajectory, and a feedforward controller keeps the aerial platform hovering stably during grasping. Practical experiments with two different robotic arms are performed. Experimental results show that the proposed modeling and control scheme provides a flexible and effective approach for an aerial manipulator with a replaceable robotic arm.

The design of impedance controllers for sloped walking with a transfemoral prosthesis is a complex control problem that generally results in numerous tuning parameters. This study proposes an easy-to-tune sloped walking control scheme. While the ankle is controlled using impedance control, the knee is controlled using a hybrid strategy of impedance control and trajectory tracking. This study derived continuous, nonlinear impedance functions for the ankle and knee joints using optimization. Principal component analysis of the impedance functions revealed trends that can be used to design impedance controllers for any given slope angle. Said trends were further used to establish a tuning regime which was subsequently tested on a transfemoral prosthesis in an emulator study. The generated gait kinematics and kinetics were found to follow the trends of healthy sloped walking data.

Image contour based vision measurement is widely applied in robot manipulation and industrial automation. It is appealing to realize object-agnostic vision system, which can be conveniently reused for various types of objects. We propose the contour primitive of interest extraction network (CPieNet) based on the one-shot learning framework. First, CPieNet is featured by that its contour primitive of interest (CPI) output, a designated regular contour part lying on a specified object, provides the essential geometric information for vision measurement. Second, CPieNet has the one-shot learning ability, utilizing a support sample to assist the perception of the novel object. To realize lower-cost training, we generate support-query sample pairs from unpaired online public images, which cover a wide range of object categories. To obtain single-pixel wide contour for precise measurement, the Gabor-filters based non-maximum suppression is designed to thin the raw contour. For the novel CPI extraction task, we built the Object Contour Primitives dataset using online public images, and the Robotic Object Contour Measurement dataset using a camera mounted on a robot. The effectiveness of the proposed methods is validated by a series of experiments.

We present a novel continuous time trajectory representation based on a Chebyshev polynomial basis, which when governed by known dynamics models, allows for full trajectory and robot dynamics estimation, particularly useful for high-performance robotics applications such as unmanned aerial vehicles. We show that we can gracefully incorporate model dynamics to our trajectory representation, within a factor-graph based framework, and leverage ideas from pseudo- spectral optimal control to parameterize the state and the control trajectories as interpolating polynomials. This allows us to perform efficient optimization at specifically chosen points derived from the theory, while recovering full trajectory estimates. Through simulated experiments we demonstrate the applicability of our representation for accurate flight dynamics estimation for multirotor aerial vehicles. The representation framework is general and can thus be applied to a multitude of high-performance applications beyond multirotor platforms.

Although deep reinforcement learning (RL) has been successfully applied to a variety of robotic control tasks, it’s still challenging to apply it to real-world tasks, due to the poor sample efficiency. Attempting to overcome this shortcoming, several works focus on reusing the collected trajectory data during the training by decomposing them into a set of policy-irrelevant discrete transitions. However, their improvements are somewhat marginal since i) the amount of the transitions is usually small, and ii) the value assignment only happens in the joint states. To address these issues, this paper introduces a concise yet powerful method to construct Continuous Transition, which exploits the trajectory information by exploiting the potential transitions along the trajectory. Specifically, we propose to synthesize new transitions for training by linearly interpolating the consecutive transitions. To keep the constructed transitions authentic, we also develop a discriminator to guide the construction process automatically. Extensive experiments demonstrate that our proposed method achieves a significant improvement in sample efficiency on various complex continuous robotic control problems in MuJoCo and outperforms the advanced model-based / model-free RL methods. The source code is available1.

We present a highly efficient approach to compute continuous shortest path vector fields on arbitrarily shaped 3D triangular meshes for robot navigation in complex real-world outdoor environments. The continuity of the vector field allows to query the shortest distance, direction and geodesic path to the goal at any point within the mesh triangles, resulting in accurate paths. In order to avoid impassable areas, our wavefront propagation method runs on a modular extendable multilayer map architecture taking different geometric cost layers into account. We describe the mathematical foundation of the geodesic distances and continuous vector field computation and demonstrate the performance in real-world and multilevel environments on our campus with a tunnel, ramps and stair- cases, and in a difficult, steep forest area with a stone quarry. For reproducibility, we provide a ready-to-use ROS software stack as well as Gazebo simulations.

We propose a new optimization-based task and motion planning (TAMP) with signal temporal logic (STL) specifications for robotic sequential manipulation such as pick-and-place tasks. Given a high-level task specification, the TAMP problem is to plan a trajectory that satisfies the specification. This is, however, a challenging problem due to the difficulty of combining continuous motion planning and discrete task specifications. The optimization-based TAMP with temporal logic specifications is a promising method, but existing works use mixed integer problems (MIP) and do not scale well. To address this issue, in our approach, a new hybrid system model without discrete variables is introduced and combined with smooth approximation methods for STL. This allows the TAMP to be formulated as a nonlinear programming problem whose computational cost is significantly less than that of MIP. Furthermore, it is also possible to deal with nonlinear dynamics and geometric constraints represented by nonlinear functions. The effectiveness of the proposed method is demonstrated with both numerical experiments and a real robot.

Effective planning in model-based reinforcement learning (MBRL) and model-predictive control (MPC) relies on the accuracy of the learned dynamics model. In many instances of MBRL and MPC, this model is assumed to be stationary and is periodically re-trained from scratch on state transition experience collected from the beginning of environment interactions. This implies that the time required to train the dynamics model - and the pause required between plan executions - grows linearly with the size of the collected experience. We argue that this is too slow for lifelong robot learning and propose HyperCRL, a method that continually learns the encountered dynamics in a sequence of tasks using task-conditional hypernetworks. Our method has three main attributes: first, it includes dynamics learning sessions that do not revisit training data from previous tasks, so it only needs to store the most recent fixed-size portion of the state transition experience; second, it uses fixed-capacity hypernetworks to represent non-stationary and task-aware dynamics; third, it outperforms existing continual learning alternatives that rely on fixed-capacity networks, and does competitively with baselines that remember an ever increasing coreset of past experience. We show that HyperCRL is effective in continual model-based reinforcement learning in robot locomotion and manipulation scenarios, such as tasks involving pushing and door opening. Our project website with videos is at this link http://rvl.cs.toronto.edu/blog/2020/hypercrl/

Humans have a remarkable ability to accurately reason about future events, including the behaviors and states of mind of other agents. Consider driving a car through a busy intersection: it is necessary to reason about the physics of the vehicle, the intentions of other drivers, and their beliefs about your own intentions. For example, if you signal a turn, another driver might yield to you; or if you enter the passing lane, another driver might decelerate to give you room to merge in front. Competent drivers must plan how they can safely react to a variety of potential future behaviors of other agents before they make their next move. This requires contingency planning: explicitly planning a set of conditional actions that depend on the stochastic outcome of future events. In this work, we develop a general-purpose contingency planner that is learned end-to-end using high-dimensional scene observations and low-dimensional behavioral observations. We use a conditional autoregressive flow model for contingency planning. We show how this model can tractably learn contingencies from behavioral observations. We developed a closed-loop control benchmark of realistic multi-agent scenarios in a driving simulator (CARLA), on which we compare our method to various noncontingent methods that reason about multi-agent future behavior, and find that our contingency planning method achieves qualitatively and quantitatively superior performance.