ICRA 2019 Papers

Recent progress of deep learning improved the capability of a robot to find a proper grasp of a novel object for different grasp modalities (e.g., pinch and suction). While these previous studies consider multiple modalities separately, several studies develop multi-modal grippers that can achieve simultaneous pinch and suction grasp (multi-modal grasp fusion) for more capable and stable object manipulation. However, the previous studies with these grippers restrict the situations: simple object geometry and uncluttered environments. To overcome these difficulties, we propose a system that consists of: 1) object-class-agnostic grasp modality detection; 2) object-class-agnostic instance segmentation; and 3) grasp template matching for different modalities. The key idea of our work is the introduction of instance segmentation to fuse multiple modalities regarding each instance eluding a grasp of multiple objects at once. In the experiments, we evaluated the proposed system on the real-world picking task in clutter. The experimental results show that the effectiveness of modality detection, instance segmentation, and the integrated system as a whole.

In this work we present a state lattice based approach for motion planning in mobile robotics. Sensing and motion uncertainty are managed at planning time to obtain safe and optimal paths. To do this reliably, our approach estimates the probability of collision taking into account the robot shape and the uncertainty in heading. We also introduce a novel graduated fidelity approach and a multi-resolution heuristic which adapt to the obstacles in the map, improving the planning efficiency while maintaining its performance. Results for different environments, shapes and motion models are reported, including experiments with real robots.

We formulate a new problem as Object Importance Estimation (OIE) in on-road driving videos, where the road users are considered as important objects if they have influence on the control decision of the ego-vehicle's driver. The importance of a road user depends on both its visual dynamics, e.g., appearance, motion and location, in the driving scene and the driving goal, e.g., the planned path, of the ego vehicle. We propose a novel framework that incorporates both visual model and goal representation to conduct OIE. To evaluate our framework, we collect an on-road driving dataset at traffic intersections in the real world and conduct human-labeled annotation of the important objects. Experimental results show that our goal-oriented method outperforms baselines and has much more improvement on the left-turn and right-turn scenarios. Furthermore, we explore the possibility of using object importance for driving control prediction and demonstrate that binary brake prediction can be improved with the information of object importance.

In this paper, we propose the reinforcement learning based multi-robot collision avoidance approach by learning collision. Dynamical path re-planning, which is massively used in classical collision avoidance methods, needs overall information of the environment. Also, training agent robots to avoid the collision and pursue a goal point simultaneously is inefficient since the agent should learn two tasks. As the number of tasks that the agent should learn increases, it is difficult to make the performance of an algorithm consistent, which is known as reproducibility issue. To overcome these limitations, Collision Avoidance by Learning Collision (CALC), which learns collision instead of avoiding an obstacle robot is suggested. To solve the collision avoidance problem efficiently, the proposed method divides the problem into training and planning. In the training algorithm, an agent robot learns how to collide with a single obstacle robot and then generates a trained policy. With the trained policy, the agent can pursue a goal point since the policy leads the agent to `collide' with the goal. Furthermore, by taking action in a reverse way from the trained policy, the agent can avoid multiple obstacle robots in the planning algorithm at once. The proposed method is validated both in the robot simulation and real robot experiment, and compared with the existing collision avoidance method.

Understanding how people are likely to behave in an environment is a key requirement for efficient and safe robot navigation. However, mobile platforms are subject to spatial and temporal constraints, meaning that only partial observations of human activities are typically available to a robot, while the activity patterns of people in a given environment may also change at different times. To address these issues we present as the main contribution an exploration strategy for acquiring models of pedestrian flows, which decides not only the locations to explore but also the times when to explore them. The approach is driven by the uncertainty from multiple Poisson processes built from past observations. The approach is evaluated using two long-term pedestrian datasets, comparing its performance against uninformed exploration strategies. The results show that when using the uncertainty in the exploration policy, model accuracy increases, enabling faster learning of human motion patterns.

This paper presents a vision-based technique and a system developed for the global reconstruction of three-dimensional (3-D) road surfaces. Using the system, the technique globally reconstructs 3-D road surfaces by estimating the global camera pose using the Adaptive Extended Kalman Filter (AEKF) and integrating it with existing local road surface reconstruction techniques. The AEKF adaptively updates the covariance of uncertainties such that the estimation works well even in environments with varying uncertainties. Numerical results show the efficacy of the proposed technique over the Extended Kalman Filter (EKF)-based technique by 50% in accuracy, and the on-road test has demonstrated the ability of the proposed technique for the real-world global 3-D road surface reconstruction.

Global localization lies at the heart of autonomous navigation and Simultaneous Localization and Mapping (SLAM). The appearance-based approach has been successful, but still faces many open challenges in environments where visual conditions vary significantly over time. In this paper, we propose an integrated solution to leverage object-level dense semantics and spatial understanding of the environment for global localization. Our approach models an environment with 3D dense semantics, semantic graph and their topology. This object-level representation is then used for place recognition via semantic object association, followed by 6-DoF pose estimation by the semantic-level point alignment. Extensive experiments show that our approach can achieve robust global localization under extreme appearance changes. It is also capable of coping with other challenging scenarios, such as dynamic environments and incomplete query observations.

A novel wearable device for gesture recognition was developed and tested on five subjects. The low-cost, wireless wearable device was engineered with a set of seven flexible capacitive touch electrodes sewn into an armband to be worn on the forearm between the wrist and elbow. These capacitive touch electrodes were interfaced with a microcontroller and bluetooth transceiver for measurement and transmission. As different gestures are made, flexing muscles beneath the skin affect the capacitance measured on these seven electrodes. A set of 32 gestures were tested including the 16 grasps in the Cutkosky Grasp Taxonomy and 16 basic finger and wrist motions. Several classification algorithms were tested on this data. Using a Random Forest (RF) algorithm to classify the training data, an average gesture recognition accuracy of 95.6 ± 0.06% was achieved across all five subjects individually.

While Product of Exponentials (POE) formula has been gaining maturity in modeling the kinematics of a serial-link robot, the Denavit-Hartenberg (D-H) notation is still the most widely used due to its intuitive and concise geometric interpretation of the robot. This paper has developed an analytical solution to automatically convert a POE model into a D-H model for a robot with revolute, prismatic, and helical joints, which are the complete set of three basic one degree of freedom lower pair joints for constructing a serial-link robot. The conversion algorithm developed can be used in applications such as calibration where it is necessary to convert the D-H model to the POE model for identification and then back to the D-H model for compensation. The equivalence of the two models proved in this paper also benefits the analysis of the identifiability of the kinematic parameters. It is found that the maximum number of identifiable parameters in a general POE model is 5h+4r+2t+n+6 where h, r, t, and n stand for the number of helical, revolute, prismatic, and general joints, respectively. It is also suggested that the identifiability of the base frame and the tool frame in the D-H model is restricted rather than the arbitrary six parameters as assumed previously.

We propose a geometric method to solve inverse kinematics (IK) problems of 7-DoF manipulators with joint offsets at shoulder, elbow, and wrist. Traditionally, inverse position kinematics for redundant manipulators are solved by using an iterative method based on the pseudo-inverse of the manipulator Jacobian. This provides a single solution among the infinitely many possible solutions for the IK problem of redundant manipulators. There are no closed-form IK solutions for redundant manipulators with multiple joint offsets. Using our method we can compute multiple IK solutions using two-parameter search by exploiting geometry of the structure of a redundant manipulator. Our proposed IK algorithm can handle multiple joint offsets and is mathematically simple to implement in a few lines of code. We apply our algorithm to compute IK solutions for 7-DoF redundant Baxter robot (that has joint offsets at shoulder, wrist, and elbow joints) for end-effector configurations where existing geometry-based IK solvers fail to find solutions. We also demonstrate the use of our algorithm in an application where we want to compute an IK solution (among the infinitely many possible solutions) that has minimum error bound in end-effector position, in the presence of random joint actuation and sensing uncertainties.

Procedural generation of elastic structures provides the fundamental basis for controlling and designing 3D printed deformable object behaviors. The automation through generative algorithms provides flexibility in how design and functionality can be seamlessly integrated into a cohesive process that generates 3D prints with variable elasticity. Generative deformation introduces an automated method for perforating existing volumetric structures, promoting simulated deformations, and integrating stress analysis into a cohesive pipeline model that can be used with existing consumer-level 3D printers with elastic material capabilities. In this work, we present a consolidated implementation of the design, simulate, refine, and 3D print procedure based on the automated generation of heterogeneous lattice structures. We utilize Finite Element Analysis (FEA) metrics to generate perforated deformation models that adhere to deformation behaviors created within our design environment. We present the core algorithms, automated pipeline, and 3D print deformations of various objects. Quantitative results illustrate how the heterogeneous geometric structure can influence elastic material behaviors towards design objectives. Our method provides an automated open-source tool for quickly prototyping elastic 3D prints.

We pose the problem of path-constrained trajectory generation for the synchronous motion of multi-robot systems as a non-linear optimization problem. Our method determines appropriate parametric representation for the configuration variables, generates an approximate solution as a starting point for the optimization method, and uses successive refinement techniques to solve the problem in a computationally efficient manner. We have demonstrated the effectiveness of the proposed method on challenging simulation and physical experiments with high degrees of freedom robotic systems.

The author has investigated the method of stealth walking for generating adaptive walking gaits of underactuated walkers without having the control torques at the feet. This approach is also effective for achieving careful walking on the frictionless road surface by applying angular momentum constraint control (AMCC); the generated gait completes in one step while maintaining the horizontal ground reaction force to zero. The result is mathematically thorough, but is not realistic because any uncertainties in the system cannot be permitted. This paper then discusses more realistic slidingresistant situation: stealth walking on the low-friction road surface. First, we introduce a model of a planar underactuated rimless wheel, and describe the equation of motion and the control input for AMCC. Second, we specify the linearized equation of motion with AMCC, and derive the analytical solution of the stance-leg motion which is used as a desired trajectory for the nonlinear model. Furthermore, we discuss the optimality of the upper-body control during the double-limb support phase from the sliding-resistant characteristics point of view through mathematical and numerical investigations.

In recent years self-driving vehicles have become more commonplace on public roads, with the promise of bringing safety and efficiency to modern transportation systems. Increasing the reliability of these vehicles on the road requires an extensive suite of software tests, ideally performed on high-fidelity simulators, where multiple vehicles and pedestrians interact with the self-driving vehicle. It is therefore of critical importance to ensure that self-driving software is assessed against a wide range of challenging simulated driving scenarios. The state of the art in driving scenario generation, as adopted by some of the front-runners of the self-driving car industry, still relies on human input [1]. In this paper we propose to automate the process using Bayesian optimization to generate adversarial self-driving scenarios that expose poorly-engineered or poorly-trained self-driving policies, and increase the risk of collision with simulated pedestrians and vehicles. We show that by incorporating the generated scenarios into the training set of the self-driving policy, and by fine-tuning the policy using vision-based imitation learning we obtain safer self-driving behavior.

In the context of imitation learning, several approaches have been developed so as to transfer human skills to robots, with demonstrations often represented in Cartesian or joint space. While learning Cartesian positions suffices for many applications, the end-effector orientation is required in many others. However, several crucial issues arising from learning orientations have not been adequately addressed yet. For instance, how can demonstrated orientations be adapted to pass through arbitrary desired points that comprise orientations and angular velocities? In this paper, we propose an approach that is capable of learning multiple orientation trajectories and adapting learned orientation skills to new situations (e.g., via-point and end-point), where both orientation and angular velocity are addressed. Specifically, we introduce a kernelized treatment to alleviate explicit basis functions when learning orientations. Several examples including comparison with the state-of-the-art dynamic movement primitives are provided to verify the effectiveness of our method.

We present a general policy formulation for partially observable Markov decision processes (POMDPs) called controller family policies that may be used as a framework to facilitate the design of new policy forms. We prove how modern approximate policy forms: point-based, finite state controller (FSC), and belief compression, are instances of this family of generalized controller policies. Our analysis provides a deeper understanding of the POMDP model and suggests novel ways to design POMDP solutions that can combine the benefits of different state-of-the-art methods. We illustrate this capability by creating a new customized POMDP policy form called the belief-integrated FSC (BI-FSC) tailored to overcome the shortcomings of a state-of-the-art algorithm that uses non-linear programming (NLP). Specifically, experiments show that for NLP the BI-FSC offers improved performance over a vanilla FSC-based policy form on benchmark domains. Furthermore, we demonstrate the BI-FSC's execution on a real robot navigating in a maze environment. Results confirm the value of using the controller family policy as a framework to design customized policies in POMDP robotic solutions.

Deep reinforcement learning provides a promising approach for vision-based control of real-world robots. However, the generalization of such models depends critically on the quantity and variety of data available for training. This data can be difficult to obtain for some types of robotic systems, such as fragile, small-scale quadrotors. Simulated rendering and physics can provide for much larger datasets, but such data is inherently of lower quality: many of the phenomena that make the real-world autonomous flight problem challenging, such as complex physics and air currents, are modeled poorly or not at all, and the systematic differences between simulation and the real world are typically impossible to eliminate. In this work, we investigate how data from both simulation and the real world can be combined in a hybrid deep reinforcement learning algorithm. Our method uses real-world data to learn about the dynamics of the system, and simulated data to learn a generalizable perception system that can enable the robot to avoid collisions using only a monocular camera. We demonstrate our approach on a real-world nano aerial vehicle collision avoidance task, showing that with only an hour of real-world data, the quadrotor can avoid collisions in new environments with various lighting conditions and geometry. Code, instructions for building the aerial vehicles, and videos of the experiments can be found at github.com/gkahn13/GtS.

Assistive robotic systems endeavour to support those with movement disabilities, enabling them to move again and regain functionality. Main issue with these systems is the complexity of their low-level control, and how to translate this to simpler, higher level commands that are easy and intuitive for a human user to interact with. We have created a multi-modal system, consisting of different sensing, decision making and actuating modalities, leading to intuitive, human-in-the-loop assistive robotics. The system takes its cue from the user's gaze, to decode their intentions and implement low-level motion actions to achieve high-level tasks. This results in the user simply having to look at the objects of interest, for the robotic system to assist them in reaching for those objects, grasping them, and using them to interact with other objects. We present our method for 3D gaze estimation, and grammars-based implementation of sequences of action with the robotic system. The 3D gaze estimation is evaluated with 8 subjects, showing an overall accuracy of 4.68\pm 0.14cm. The full system is tested with 5 subjects, showing successful implementation of 100% of reach to gaze point actions and full implementation of pick and place tasks in 96%, and pick and pour tasks in 76% of cases. Finally we present a discussion on our results and what future work is needed to improve the system.

Control of underactuated balance robot requires external subsystem trajectory tracking and internal unstable subsystem balancing with limited control authority. We present a learning-based control approach for underactuated balance robots. The tracking and balancing control is designed the controller in fast- and slow-time scales. In the slow-time scale, model predictive control is adopted to plan desired internal state profile to achieve external trajectory tracking task. The internal state is then stabilized around the planned profile in the fast-time scale. The control design is based on a learned Gaussian process (GP) regression model without need of a priori knowledge about the robot dynamics. The controller also incorporates the GP model predicted variance to enhance robustness to modeling errors. Experiments are presented using a Furuta pendulum system.