GHS Robotics

We consider the problem of grasping novel objects, specifically objects that are being seen for the first time through vision. Grasping a previously unknown object, one for which a 3-d model is not available, is a challenging problem. Furthermore, even if given a model, one still has to decide where to grasp the object. We present a learning algorithm that neither requires nor tries to build a 3-d model of the object. Given two (or more) images of an object, our algorithm attempts to identify a few points in each image corresponding to good locations at which to grasp the object. This sparse set of points is then triangulated to obtain a 3-d location at which to attempt a grasp. This is in contrast to standard dense stereo, which tries to triangulate every single point in an image (and often fails to return a good 3-d model). Our algorithm for identifying grasp locations from an image is trained by means of supervised learning, using synthetic images for the training set. We demonstrate this approach on two robotic manipulation platforms. Our algorithm successfully grasps a wide variety of objects, such as plates, tape rolls, jugs, cellphones, keys, screwdrivers, staplers, a thick coil of wire, a strangely shaped power horn and others, none of which were seen in the training set. We also apply our method to the task of unloading items from dishwashers.

Automatically building maps from sensor data is a necessary and fundamental skill for mobile robots; as a result, considerable research attention has focused on the technical challenges inherent in the mapping problem. While statistical inference techniques have led to computationally efficient mapping algorithms, the next major challenge in robotic mapping is to automate the data collection process. In this paper, we address the problem of how a robot should plan to explore an unknown environment and collect data in order to maximize the accuracy of the resulting map. We formulate exploration as a constrained optimization problem and use reinforcement learning to find trajectories that lead to accurate maps. We demonstrate this process in simulation and show that the learned policy not only results in improved map building, but that the learned policy also transfers successfully to a real robot exploring on MIT campus.

One of the most general frameworks for phrasing control problems for complex, redundant robots is operational-space control. However, while this framework is of essential importance for robotics and well understood from an analytical point of view, it can be prohibitively hard to achieve accurate control in the face of modeling errors, which are inevitable in complex robots (e.g. humanoid robots). In this paper, we suggest a learning approach for operational-space control as a direct inverse model learning problem. A first important insight for this paper is that a physically correct solution to the inverse problem with redundant degrees of freedom does exist when learning of the inverse map is performed in a suitable piecewise linear way. The second crucial component of our work is based on the insight that many operational-space controllers can be understood in terms of a constrained optimal control problem. The cost function associated with this optimal control problem allows us to formulate a learning algorithm that automatically synthesizes a globally consistent desired resolution of redundancy while learning the operational-space controller. From the machine learning point of view, this learning problem corresponds to a reinforcement learning problem that maximizes an immediate reward. We employ an expectation-maximization policy search algorithm in order to solve this problem. Evaluations on a three degrees-of-freedom robot arm are used to illustrate the suggested approach. The application to a physically realistic simulator of the anthropomorphic SARCOS Master arm demonstrates feasibility for complex high degree-of-freedom robots. We also show that the proposed method works in the setting of learning resolved motion rate control on a real, physical Mitsubishi PA-10 medical robotics arm.

In this paper we describe a learning framework for a central pattern generator (CPG)-based biped locomotion controller using a policy gradient method. Our goals in this study are to achieve CPG-based biped walking with a 3D hardware humanoid and to develop an efficient learning algorithm with CPG by reducing the dimensionality of the state space used for learning. We demonstrate that an appropriate feedback controller can be acquired within a few thousand trials by numerical simulations and the controller obtained in numerical simulation achieves stable walking with a physical robot in the real world. Numerical simulations and hardware experiments evaluate the walking velocity and stability. The results suggest that the learning algorithm is capable of adapting to environmental changes. Furthermore, we present an online learning scheme with an initial policy for a hardware robot to improve the controller within 200 iterations.

This paper presents the design, control and performance of a high fidelity four degree-of-freedom wrist exoskeleton robot, RiceWrist, for training and rehabilitation. The RiceWrist is intended to provide kinesthetic feedback during the training of motor skills or rehabilitation of reaching movements. Motivation for such applications is based on findings that show robot-assisted physical therapy aids in the rehabilitation process following neurological injuries. The exoskeleton device accommodates forearm supination and pronation, wrist flexion and extension and radial and ulnar deviation in a compact parallel mechanism design with low friction, zero backlash and high stiffness. As compared to other exoskeleton devices, the RiceWrist allows easy measurement of human joint angles and independent kinesthetic feedback to individual human joints. In this paper, joint-space as well as task-space position controllers and an impedance-based force controller for the device are presented. The kinematic performance of the device is characterized in terms of its workspace, singularities, manipulability, backlash and backdrivability. The dynamic performance of RiceWrist is characterized in terms of motor torque output, joint friction, step responses, behavior under closed loop set-point and trajectory tracking control and display of virtual walls. The device is singularity-free, encompasses most of the natural workspace of the human joints and exhibits low friction, zero-backlash and high manipulability, which are kinematic properties that characterize a high-quality impedance display device. In addition, the device displays fast, accurate response under position control that matches human actuation bandwidth and the capability to display sufficiently hard contact with little coupling between controlled degrees-of-freedom.

This paper presents a user-centered, model-mediated approach to bilateral telemanipulation under large communication delays. Assuming that the environment is only slowly or infrequently changing, it mitigates user perception difficulties and allows stable motion and force interactions with the remote environment. Rather than directly sending slave sensory data to the user, the method abstracts the data to form a very simple model of the environment. The model is transmitted to the master, where it is haptically rendered for user feedback without any lag. In return, the slave only executes force or motion commands consistent with the model. This effectively mediates the master—slave interaction. Particular attention is placed on the system behavior as it adjusts to unexpected environment interactions and as the model updates. The basic principles of the approach are demonstrated on a simple one degree-of-freedom telerobotic system operating with four seconds of round-trip delay.

The paper describes the design and control of a transfemoral prosthesis with powered knee and ankle joints. The initial prototype is a pneumatically actuated powered-tethered device, which is intended to serve as a laboratory test bed for a subsequent self-powered version. The prosthesis design is described, including its kinematic optimization and the design of a three-axis socket load cell that measures the forces and moments of interaction between the socket and prosthesis. A gait controller is proposed based on the use of passive impedance functions that coordinates the motion of the prosthesis with the user during level walking. The control approach is implemented on the prosthesis prototype and experimental results are shown that demonstrate the promise of the active prosthesis and control approach in restoring fully powered level walking to the user.

Robots must complete their tasks in spite of unreliable actuators and limited, noisy sensing. In this paper, we consider the information requirements of such tasks. What sensing and actuation abilities are needed to complete a given task? Are some robot systems provably "more powerful", in terms of the tasks that they can complete, than others? Can we find meaningful equivalence classes of robot systems? This line of research is inspired by the theory of computation, which has produced similar results for abstract computing machines. Our basic contribution is a dominance relation over robot systems that formalizes the idea that some robots are stronger than others. This comparison, which is based on how the robots progress through their information spaces, induces a partial order over the set of robot systems. We prove some basic properties of this partial order and show that it is directly related to the robots' ability to complete tasks. We give examples to demonstrate the theory, including a detailed analysis of a limited-sensing global localization problem.

Ultrasound image-guided interventions are widespread in surgery because of the non-invasive character of the procedures. However, hand/eye synchronization is relatively difficult for a surgeon. Ultrasound image-based visual servoing is one way to perform this kind of surgery. In this work, the control of instrument motion based on ultrasound images through nonlinear model predictive control is investigated. This new scheme ensures the convergence of the instrument to the desired position and also offers the possibility of satisfying constraints such as joint limits, actuator saturation and visibility preserving. This paper describes the proposed controller. The efficiency and the robustness of the proposed solution to control a six degree-of-freedom mechanical system is first illustrated by simulation. Experiments on a Mitsubishi PA10 robot highlight the efficiency of the vision control scheme to handle constraints of ultrasound image-based visual servoing.

This paper presents an image-based visual servoing strategy for the autonomous navigation of a mobile holonomic robot from a current towards a desired pose, specified only through a current and a desired image acquired by the on-board central catadioptric camera. This kind of vision sensor combines lenses and mirrors to enlarge the field of view. The proposed visual servoing does not require any metrical information about the three-dimensional viewed scene and is mainly based on a novel geometrical property, the auto-epipolar condition, which occurs when two catadioptric views (current and desired) undergo a pure translation. This condition can be detected in real time in the image domain by observing when a set of so-called disparity conics have a common intersection. The auto-epipolar condition and the pixel distances between the current and target image features are used to design the image-based control law. Lyapunov-based stability analysis and simulation results demonstrate the parametric robustness of the proposed method. Experimental results are presented to show the applicability of our visual servoing in a real context.

Transparent teleoperation under rate mode has proven to be difficult in terms of stability, performance and implementation. This is mainly due to the need for an exchange of derivative and integral of measured positions and forces which make transparent rate mode controllers prone to noise and abrupt contact force changes. Moreover, the performance of controllers declines in the presence of communication delays. This paper proposes two control architectures based on the use of local force feedback (LFF) and environment impedance reflection (EIR). The LFF controllers eliminate a force channel while preserving transparency under ideal conditions. In the EIR controller, the identified impedance of the environment is employed in the master controller to predict the slave position and contact force derivatives. The stability robustness and performance of these controllers are evaluated and compared to those of a benchmark controller under different operational conditions, such as noise and delay, using analytical methods and experimental results.

Rao—Blackwellized particle filters (RBPFs) are an implementation of sequential Bayesian filtering that has been successfully applied to mobile robot simultaneous localization and mapping (SLAM) and exploration. Measuring the uncertainty of the distribution estimated by a RBPF is required for tasks such as information gain-guided exploration or detecting loop closures in nested loop environments. In this paper we propose a new measure that takes the uncertainty in both the robot path and the map into account. Our approach relies on the entropy of the expected map (EM) of the RBPF, a new variable built by integrating the map hypotheses from all of the particles. Unlike previous works that use the joint entropy of the RBPF for active exploration, our proposal is better suited to detect opportunities to close loops, a key aspect to reduce the robot path uncertainty and consequently to improve the quality of the maps being built. We provide a theoretical discussion and experimental results with real data that support our claims.

Autonomous Vehicle Following can be achieved if the poses of both the follower and leader vehicles are continuously estimated. This can be achieved by using a Bayesian estimation technique together with a virtual trailer link model. The advantage of such a model is that the follower vehicle will trail a virtual trailer, modeled as an attachment to the leader vehicle, instead of the leader vehicle itself, so that a safe spacing between the two vehicles is guaranteed. The key to a tractable solution to this vehicle following problem is the justifiable assumption that the pose of the follower vehicle is statistically independent of that of the leader. This assumption is valid when conditioned on the history of the follower vehicle's inputs and the sensor observations made by the follower vehicle. Hence, a factored solution to the joint estimate of the follower and leader poses can be formulated. Due to the factored solution, the pose of the follower vehicle is estimated separately using a recursive estimator. In a separate estimator, the poses of the virtual tailer and the leader vehicle are augmented in the tracking process of the leader vehicle. The aim is to command the follower vehicle to trail the estimated pose of the virtual trailer link model. The pose of the virtual trailer is computed with an on-board sensor mounted on the follower vehicle. A case study on the implementation of the proposed formulation, using an Extended Kalman Filter as the main estimator, is presented. First, simulation results are presented. To make the simulation results comparable to the actual system, the variances of sensor measurements are set according to real sensor data-sheet values. Various types of vehicle maneuver, such as straight paths and clothoids with left and right transition paths, are considered. Real experiments are also carried out in a car park. A comparison of the estimated paths and the best available ground truth is presented. The path deviations of the proposed system are also compared with similar systems in the literature.

In this paper we present a theoretical and experimental result on the control of multi-agent non-holonomic systems. We design and implement a novel decentralized control scheme that achieves dynamic formation control and collision avoidance for a group of non-holonomic robots. First, we derive a feedback law using Lyapunov-type analysis that guarantees collision avoidance and tracking of a reference trajectory for a single robot. Then we extend this result to the case of multiple non-holonomic robots, and show how different multi-agent problems, such as formation control and leader—follower control, can be addressed in this framework. Finally, we combine the above results to address the problem of coordinated tracking for a group of agents. We give extensive experimental results that validate the effectiveness of our results in all three cases.

In the world of living creatures, simple-minded animals often cooperate to achieve common goals with amazing performance. One can consider this idea in the context of robotics, and suggest models for programming goal-oriented behavior into the members of a group of simple robots lacking global supervision. This can be done by controlling the local interactions between the robot agents, to have them jointly carry out a given mission. As a test case we analyze the problem of many simple robots cooperating to clean the dirty floor of a non-convex region in Z2, using the dirt on the floor as the main means of inter-robot communication.

While there are many benefits to minimally invasive surgery (MIS), force feedback or touch sensation is limited in the currently available MIS tools, such as surgical robots, creating the potential for excessive force application during surgery and unintended tissue injury. The goal of this work was to develop a methodology with which to identify stress magnitudes and durations that can be safely applied with a MIS grasper to di ferent tissues, potentially improving MIS device design and reducing potentially adverse clinically relevant consequences. Using the porcine model, stresses typically applied in MIS were applied to liver, ureter and small bowel using a motorized endoscopic grasper. Acute indicators of tissue damage including cellular death and infiltration of inflammatory cells were measured using histological and image analysis techniques. Finite element analysis was used to identify approximate stress distributions experienced by the tissues. Parameters used in these finite element models specifically reflected the properties of liver, which served as an initial proxy for all tissues, as stress distributions rather than absolute values were desired. Local regions predicted to have uniform stress by the computational models were mapped to and analyzed in the tissue samples for acute damage. Analysis of variance (ANOVA) and post-hoc analyses were used to detect stress magnitudes and durations that caused significantly increased tissue damage with the goal to ultimately identify safe stress `thresholds' during grasping of the studied tissues. Preliminary data suggests a graded non-linear response between applied stress magnitude and apoptosis in liver and small bowel as well as neutrophil infiltration in the small bowel. The ureter appeared to be more resistant to injury at the tested stress levels. By identifying stress magnitudes and durations within the range of grasping loads applied in MIS, it may be possible for researchers to create a `smart' surgical robot that can guide a surgeon to manipulate tissues with minimal resulting damage. In addition, surgical simulator design can be improved to reflect more realistic tissue responses and evaluate trainees' tissue handling skills.