RSS 2017 Papers

This paper presents a framework for optimizing both the shape and the motion of a planar rigid end-effector to satisfy a desired manipulation task. We frame this design problem as a nonlinear optimization program, where shape and motion are decision variables represented as splines. The task is represented as a series of constraints, along with a fitness metric,which force the solution to be compatible with the dynamics of frictional hard contact while satisfying the task. We illustrate the approach with the example problem of moving a disk along a desired path or trajectory, and we verify it by applying it to three classical design problems: the rolling brachistochrone, the design of teeth of involute gears, and the pitch curve of rolling cams. We conclude with a case study involving the optimization and real implementation of the shape and motion of a dynamic throwing arm.

This paper explores the application of Koopman operator theory to the control of robotic systems. The operator is introduced as a method to generate data-driven models that have utility for model-based control methods. We then motivate the use of the Koopman operator towards augmenting model-based control. Specifically, we illustrate how the operator can be used to obtain a linearizable data-driven model for an unknown dynamical process that is useful for model-based control synthesis. Simulated results show that with increasing complexity in the choice of the basis functions, a closed-loop controller is able to invert and stabilize a cart- and VTOL-pendulum systems. Furthermore, the specification of the basis function are shown to be of importance when generating a Koopman operator for specific robotic systems. Experimental results with the Sphero SPRK robot explore the utility of the Koopman operator in a reduced state representation setting where increased complexity in the basis function improve open- and closed-loop controller performance in various terrains, including sand.

In this paper, we develop an algorithm for intent inference via goal disambiguation with a shared-control assistive robotic arm. Assistive systems are often required to infer human intent and this often is a bottleneck for providing assistance quickly and accurately. We introduce the notion of inverse legibility in which the human-generated actions are legible enough for the robot to infer the human intent confidently and accurately. The proposed disambiguation paradigm seeks to elicit legible control commands from the human by selecting control modes, for the robotic arm, in which human-directed motion will maximally disambiguate between multiple goals. We present simulation results which look into the robustness of our algorithm and the impact of the choice of confidence functions on the performance of the system. Our simulations results suggest that the choice of confidence function is a critical factor in determining the disambiguation algorithm's capability to capture human intent. We also present a pilot study that explores the efficacy of the algorithm on real hardware with promising preliminary results.

Real world scenarios contain many structural patterns that, if appropriately extracted and modeled, can be used to reduce problems associated with sensor failure and occlusions, while improving planning methods in tasks such as navigation and grasping. This paper devises a novel unsupervised procedure that is able to learn 3D structures from unorganized point clouds as occupancy maps. Our framework enables the learning of unique and arbitrarily complex features using a Bayesian Convolutional Variational Auto-Encoder that compresses local information into a latent low-dimensional representation and then decodes it back in order to reconstruct the original scene. This reconstructive model is trained on features obtained automatically from a wide variety of scenarios to improve its generalization and interpolative powers. We show that the proposed framework is able to recover partially missing structures and reason over occlusion with high accuracy, while maintaining a detailed reconstruction of observed areas. To seamlessly combine this localized feature information into a single global structure, we employ a Hilbert Map, recently proposed as a robust and efficient occupancy mapping technique. Experimental tests are conducted in large-scale 2D and 3D datasets, and a study on the impact of various accuracy/speed trade-offs is provided to assess the limits of the proposed framework.

We present a novel approach to shared control of human-machine systems. Our method assumes no a priori knowledge of the system dynamics. Instead, we learn both the dynamics and information about the user's interaction from observation through the use of the Koopman operator. Using the learned model, we define an optimization problem to compute the optimal policy for a given task, and compare the user input to the optimal input. We demonstrate the efficacy of our approach with a user study. We also analyze the individual nature of the learned models by comparing the effectiveness of our approach when the demonstration data comes from a user's own interactions, from the interactions of a group of users and from a domain expert. Positive results include statistically significant improvements on task metrics when comparing a user-only control paradigm with our shared control paradigm. Surprising results include findings that suggest that individualizing the model based on a user's own data does not effect the ability to learn a useful dynamic system. We explore this tension as it relates to developing human-in-the-loop systems further in the discussion.

We present a technique for learning control Lyapunov (potential) functions, which are used in turn to synthesize controllers for nonlinear dynamical systems. The learning framework uses a demonstrator that implements a black-box, untrusted strategy presumed to solve the problem of interest, a learner that poses finitely many queries to the demonstrator to infer a candidate function and a verifier that checks whether the current candidate is a valid control Lyapunov function. The overall learning framework is iterative, eliminating a set of candidates on each iteration using the counterexamples discovered by the verifier and the demonstrations over these counterexamples. We prove its convergence using ellipsoidal approximation techniques from convex optimization. We also implement this scheme using nonlinear MPC controllers to serve as demonstrators for a set of state and trajectory stabilization problems for nonlinear dynamical systems. Our approach is able to synthesize relatively simple polynomial control Lyapunov functions, and in that process replace the MPC using a guaranteed and computationally less expensive controller.

Relying on reduced models is nowadays a standard cunning to tackle the computational complexity of multi-contact locomotion. To be really effective, reduced models must respect some feasibility constraints in regards to the full model. However, such kind of constraints are either partially considered or just neglected inside the existing reduced problem formulation. This work presents a systematic approach to incorporate feasibility constraints inside trajectory optimization problems. In particular, we show how to learn the kinematic feasibility of the centre of mass to be achievable by the whole-body model. We validate the proposed method in the context of multi-contact locomotion: we perform two stairs climbing experiments on two humanoid robots, namely the HRP-2 robot and the new TALOS platform.

Many practical tasks in robotic systems, such as cleaning windows, writing or grasping, are inherently constrained. Learning policies subject to constraints is a challenging problem. We propose a \emph{locally weighted constrained projection learning} method (LWCPL) that first estimates the constraint and then exploits this estimate across multiple observations of the constrained motion to learn an unconstrained policy. The generalization is achieved by projecting the unconstrained policy onto a new, previously unseen, constraint. We do not require any prior knowledge about the task or the policy, so we can use generic regressors to model the task and the policy. However, any prior beliefs about the structure of the motion can be expressed by choosing task-specific regressors. In particular, we can use robot kinematics and motion priors to improve the accuracy. Our evaluation results show that LWCPL outperform the state of the art method in accuracy of learning the constraints as well as the unconstrained policy, even in noisy conditions. We have validated our method by learning a wiping task from human demonstration on flat surfaces and reproducing it on an unknown curved surface using a force/torque based controller to achieve tool alignment. We show that, despite of the differences between the training and validation scenarios, we learn a policy that still provides the desired wiping motion.

We present a novel computational approach to optimizing the morphological design of robotic devices. Our framework takes as input a parameterized robot design, and a motion plan consisting of end-effector trajectories and/or a body trajectory. The algorithm we propose is used to optimize a set of design parameters including link lengths and actuator layout, while concurrently adjusting motion parameters such as joint trajectories, actuator forces, and contact forces. Our key insight is that the complex relationship between design and motion parameters can be made explicit through the implicit function theorem if movements are modeled as spatio-temporal solutions to optimal control problems. This explicitly modeled relationship allows us to formulate the task of optimizing robot designs using quadratic programming. We evaluate the model we propose by optimizing two simulated robots that employ linear actuators: a manipulator and a large quadruped. We further validate our framework by optimizing the design of a small quadrupedal robot and testing its performance using a hardware implementation.

This article analyzes two classes of job selection policies that control how a network of autonomous aerial vehicles delivers goods from depots to customers. Customer requests (jobs) occur according to a spatio-temporal stochastic process not known by the system. If job selection uses a policy in which the first job (FJ) is served first, the system may collapse to instability by removing just one vehicle. Policies that serve the nearest job (NJ) first show such threshold behavior only in some settings and can be implemented in a distributed manner. The timing of job selection has significant impact on delivery time and stability for NJ while it has no impact for FJ. Based on these findings we introduce a methodological approach for decision- making support to set up and operate such a system, taking into account the trade-off between monetary cost and service quality. In particular, we compute a lower bound for the infrastructure expenditure required to achieve a certain expected delivery time. The approach includes three time horizons: long-term decisions on the number of depots to deploy in the service area, mid- term decisions on the number of vehicles to use, and short-term decisions on the policy to operate the vehicles.

Modern perception systems are notoriously complex, featuring dozens of interacting parameters that must be tuned to achieve good performance. Conventional tuning approaches require expensive ground truth, while heuristic methods are difficult to generalize. In this work, we propose an introspective ground-truth-free approach to evaluating the performance of a generic perception system. By using the posterior distribution estimate generated by a Bayesian estimator, we show that the expected performance can be estimated efficiently and without ground truth. Our simulated and physical experiments in a demonstrative indoor ground robot state estimation application show that our approach can order parameters similarly to using a ground-truth system, and is able to accurately identify top-performing parameters in varying contexts. In contrast, baseline approaches that reason only about observation log-likelihood fail in the face of challenging perceptual phenomena.

We present a motion planning algorithm to compute collision-free and smooth trajectories for high-DOF robots interacting with humans in a shared workspace. Our approach uses offline learning of human actions along with temporal coherence to predict the human actions. Our intention-aware online planning algorithm uses the learned database to compute a reliable trajectory based on the predicted actions. We represent the predicted human motion using a Gaussian distribution and compute tight upper bounds on collision probabilities for safe motion planning. We highlight the performance of our planning algorithm in complex simulated scenarios and real world benchmarks with 7-DOF robot arms operating in a workspace with a human performing complex tasks. We demonstrate the benefits of our intention-aware planner in terms of computing safe trajectories in such uncertain environments.

Rovers operating on Mars have been delayed, diverted, and trapped by loose granular materials. Vision-based mobility prediction approaches often fail because hazardous sand is difficult to distinguish from safe sand based on surface appearance alone. Unlike surface appearance, the thermal inertia of terrain is directly correlated to the same geophysical properties that control slip. This paper presents a quantitative analysis showing that considering thermal inertia improves rover slip prediction on Mars using in-situ data from the Curiosity rover. Thermal inertia is estimated for each slip measurement in sand using both on-board and orbital instruments. Slip models are learned using a mixture of experts approach where the experts are identified using thermal inertia. Two-expert models are compared to a single-expert, vision-only model to show that slip predictions are improved by separating high-slip, low thermal inertia sand from low-slip, high thermal inertia sand. These results support the hypothesis that the consideration of thermal inertia improves mobility estimates for rovers on Mars.

In this paper, we present an algorithm for planning and control of legged robot locomotion. Given the desired contact sequence, this method generates gaits and dynamic motions for legged robots without resorting to simplified stability criteria. The method uses direct collocation for searching for solutions within the constraint consistent subspace defined by the robot's contact configuration. For the differential equation constraints of the collocation algorithm, we use the so-called direct dynamics of a constrained multibody system. The dynamics of a legged robot is different for each contact configuration. Our method deals with such a hybrid nature, and it allows for velocity discontinuities when contacts are made. We introduce the projected impact dynamics constraint to enforce consistency during mode switching. We stabilize the plan using an inverse dynamics controller compatible with the optimal feed-forward control of the motion plan. As a whole, this approach reduces the complexity associated with specifying dynamic motions of a floating-base robot under the constant influence of contact forces. We apply this method on a hydraulically-actuated quadruped robot. We show two types of gaits (walking and trotting) as well as diverse jumping motions (forward, sideways, turning) on the real system. The results presented here are one of the few examples of an optimal control problem solved and satisfactorily transferred to a real torque-controlled legged robot.

This paper studies the underlying combinatorial structure of a class of object rearrangement problems, which appear frequently in applications. The problems involve multiple, similar-geometry objects placed on a flat, horizontal surface, where a robot can approach them from above and perform pick-and-place operations to rearrange them. The paper considers both the case where the start and goal object poses overlap, and where they do not. For overlapping poses, the primary objective is to minimize the number of pick-and-place actions and then to minimize the distance traveled by the end-effector. For the non-overlapping case, the objective is solely to minimize the end-effector distance. While such problems do not involve all the complexities of general rearrangement, they remain computationally hard challenges in both cases. This is shown through two-way reductions between well-understood, hard combinatorial challenges and these rearrangement problems. The benefit of the reduction is that there are well studied algorithms for solving these well-established combinatorial challenges. These algorithms can be very efficient in practice despite the hardness results. The paper builds on these reduction results to propose an algorithmic pipeline for dealing with the rearrangement problems. Experimental evaluation shows that the proposed pipeline achieves high-quality paths with regards to the optimization objectives. Furthermore, it exhibits highly desirable scalability as the number of objects increases in both the overlapping and non-overlapping setups.

In this paper we present a system for the state estimation of a dynamically walking and trotting quadruped. The approach fuses four heterogeneous sensor sources (inertial, kinematic, stereo vision and LIDAR) to maintain an accurate and consistent estimate of the robot's base link velocity and position in the presence of disturbances such as slips and missteps. We demonstrate the performance of our system, which is robust to changes in the structure and lighting of the environment, as well as the terrain over which the robot crosses. Our approach builds upon a modular inertial-driven Extended Kalman Filter which incorporates a rugged, probabilistic leg odometry component with additional inputs from stereo visual odometry and LIDAR registration. The simultaneous use of both stereo vision and LIDAR helps combat operational issues which occur in real applications. To the best of our knowledge, this paper is the first to discuss the complexity of consistent estimation of pose and velocity states, as well as the fusion of multiple exteroceptive signal sources at largely different frequencies and latencies, in a manner which is acceptable for a quadruped's feedback controller. A substantial experimental evaluation demonstrates the robustness and accuracy of our system, achieving continuously accurate localization and drift per distance traveled below 1 cm/m.

In this paper, we propose a solution to the problem of {\it herding by caging}: given a set of mobile robots (called herders) and a group of moving agents (called sheep), we move the latter to some predefined location in such a way that they cannot escape from the robots while moving. We model the interaction between the herders and the sheep by assuming that the former exert virtual ``repulsive forces" pushing the sheep away from them. These forces induce a potential field, in which the sheep move in a way that does not increase their potential. This enables the robots to partially control the motion of the sheep. We formalize this behavior geometrically by applying the notion of {\it caging}, widely used in robotic grasping. We show that our approach is provably correct in the sense that the sheep cannot escape from the robots. We propose an RRT-based motion planning algorithm, demonstrate its probabilistic completeness, and evaluate it in simulations.

Although groups of robots are expected to interact with groups of humans in the near future, research related to teams of humans and robots still appears scarce. This paper contributes to the study of human-robot teams by investigating how humans choose robots to partner with in a multi-party game context. The novelty of our work concerns the successful design and development of two social robots that are able to autonomously interact with a group of two humans in the execution of a social and entertaining task. The development of these two characters was motivated by psychological research on learning goal theory, according to which we interpret and approach a given task differently depending on our learning goal (oriented more towards either relationship building or competition). Thus, we developed two robotic characters implemented in two robots: Emys (competitive robot) and Glin (relationship-driven robot). In our study, a group of four (two humans and two autonomous robots) engaged in a social and entertaining card game. Our study yields several important conclusions regarding groups of humans and robots. (1) When a partner is chosen without previous partnering experience, people tend to prefer robots with relationship-driven characteristics as their partners compared with competitive robots. (2) After some partnering experience has been gained, the choice becomes less clear and additional driving factors emerge: (2a) participants with higher levels of competitiveness (personal characteristics) tend to prefer Emys, whereas those with lower levels prefer Glin, and (2b) the choice of which robot to partner with also depends on team performance, with the winning team being the preferred choice.

Motivated by applications in robotics and computer vision, we study problems related to spatial reasoning of a 3D environment using sublevel sets of polynomials. These include: tightly containing a cloud of points (e.g., representing an obstacle) with convex or nearly-convex basic semialgebraic sets, computation of Euclidean distance between two such sets, separation of two convex basic semalgebraic sets that overlap, and tight containment of the union of several basic semialgebraic sets with a single convex one. We use algebraic techniques from sum of squares optimization that reduce all these tasks to semidefinite programs of small size and present numerical experiments in realistic scenarios.