Contact-rich assembly tasks may result in large and unpredictable forces and torques when the locations of the contacting parts are uncertain. The ability to correct the trajectory in response to haptic feedback and accomplish the task despite location uncertainties is an important skill. We hypothesize that this skill would facilitate generalization and support direct transfer from simulations to real world. To reduce sample complexity, we propose to learn a residual admittance policy (RAP). RAP is learned to correct the movements generated by a baseline policy in the framework of dynamic movement primitives. Given the reference trajectories generated by the baseline policy, the action space of RAP is limited to the admittance parameters. Using deep reinforcement learning, a deep neural network is trained to map task specifications to proper admittance parameters. We demonstrate that RAP handles uncertainties in board location, generalizes well over space, size and shape, and facilitates quick transfer learning. Most impressively, we demonstrate that the policy learned in simulations achieves similar robustness to uncertainties, generalization and performance when deployed on an industrial robot (UR5e) without further training. See accompanying video for demonstrations.

Natural dynamic controllers aim to perform the desired task by exploiting the natural dynamics of the system. This can be accomplished by generating torque patterns to actuate the system rather than accurately following a predefined trajectory. We have previously demonstrated natural dynamic control of compass-biped in simulations. Here we demonstrate successful implementation of this dynamic controller on an under-actuated compass-biped robot. The parameters of the controller, and in particular the magnitude and timing of torque primitives, were optimized using multi-objective optimization via genetic algorithm, accounting for speed and energy efficiency. While the current implementation is in open-loop, this strategy can be extended to include feedback to enhance walking over a wide range of terrains. This proof-of concept provides the basis for future extensions to more complex robots.

Brain Machine Interfaces (BMIs) rely on the correlation between neural activity and movement kinematics. However, many characteristics of the neural activity change after switching from pole control to brain control. Of particular interest are changes in the preferred direction (PD), and whether they reflect adaptation to the BMI filter. Here we investigate changes in the PD of simulated neurons that encode signals that are relevant for state estimation and control with the framework of optimal feedback control (OFC). Simulated BMI experiments based on the OFC framework demonstrate that changes in the PD may occur even with no adaptation. Further theoretical and simulations indicate the conditions under which there is no change in the PD upon switching to brain control.. Insights gained from this research can be used to improve the design of BMI filter - not only to minimize reconstruction error during pole control, but also to endow the neurons with desired PDs in brain control.

This note presents a unified approach to nominal control generation and return map parameterization for yo-yo playing. The approach is based on choosing an intermediate state of the yo-yo as the parameter and formulating an optimization problem with appropriate boundary conditions that incorporate the corresponding parameter. By solving the optimization problems, both the return map and the nominal optimal control for the robot can be obtained. The return map is naturally parameterized with the intermediate state, which can be treated as a virtual control. The return map facilitates the application of standard analysis and design methods for discrete-time control systems. The approach is promising as it might be extended to control other cyclic dynamic systems.

Linear regression with high uncertainties in the measurements, model structure and model permanence is a major challenging problem. Standard regression techniques are based on optimizing a certain performance criterion, usually the mean squared error, and are highly sensitive to uncertainties. Regularization methods have been developed to address the problem of measurement uncertainty, but choosing the regularization parameter under severe uncertainties is problematic. Here we develop an alternative regression methodology based on satisfying rather than optimizing the performance criterion while maximizing the robustness to uncertainties. Uncertainties are represented by info-gap models which entail an unbounded family of nested sets of measurements parameterized by a non-probabilistic horizon of uncertainty. We prove and demonstrate that the robust-satisfying solution is different from the optimal least squares solution and that the info-gap approach can provide higher robustness to uncertainty

Rhythmic active touch, such as whisking, evokes a periodic reference spike train along which the timing of a novel stimulus, induced, for example, when the whiskers hit an external object, can be interpreted. Previous work supports the hypothesis that the whisking-induced spike train entrains a neural implementation of a phase-locked loop (NPLL) in the vibrissal system. Here we extend this work and explore how the entrained NPLL decodes the delay of the novel, contact-induced stimulus and facilitates object localization. We consider two implementations of NPLLs, which are based on a single neuron or a neural circuit, respectively, and evaluate the resulting temporal decoding capabilities. Depending on the structure of the NPLL, it can lock in either a phase- or co-phase-sensitive mode, which is sensitive to the timing of the input with respect to the beginning of either the current or the next cycle, respectively. The co-phase-sensitive mode is shown to be unique to circuit-based NPLLs. Concentrating on temporal decoding in the vibrissal system of rats, we conclude that both the nature of the information processing task and the response characteristics suggest that the computation is sensitive to the co-phase. Consequently, we suggest that the underlying thalamocortical loop should implement a circuit-based NPLL.

Robotic yoyo playing is a challenging, open-loop unstable game, which requires dynamic dexterity to stabilize. The switching control strategy proposed here stabilizes the yoyo by determining when to start its activation. A straightforward implementation is presented and demonstrated, using continuous state estimation provided by visual feedback and the previously developed dynamic model. The discrete return map associated with the original continuous system is analyzed and shown to possess a single physically relevant fixed point. Stability is guaranteed as long as the control parameter, which determines the magnitude of the acceleration, is above a certain level. The stabilizing power of the proposed control algorithm is successfully demonstrated on a real-time robotic yoyo playing system. Theoretical predictions regarding the fixed point of the return map are confirmed experimentally, providing further support for our modeling and analysis.

Different networks of coupled oscillators were developed for open-loop control of periodic motion. However, some tasks, like yo-yo playing, are open-loop unstable and require proper phase locking to stabilize. Given the phase-locking property of coupled oscillators, we investigate their application to closed-loop control of open-loop unstable systems, concentrating on the challenging task of yo-yo control. In particular, we focus on pulse-coupling, where the yo-yo sends a feedback upon reaching the bottom of the string and the onset of the oscillatory cycle is used to trigger the movement. Four networks involving either a stand-alone or a circuit level oscillator with either excitatory or inhibitory couplings are considered. Working curve analysis indicates that three of the networks cannot stabilize the yo-yo. The fourth network, which is based on a circuit-level oscillator, is analyzed using the return map and the region of stability is determined and verified by simulations. The resulting pulse-coupled oscillatory control provides a model-free control strategy that operates with an easy-to-measure low-rate feedback.

We propose a new method for implementing impedance control, which is designed to take advantage of the error-correction capabilities of position controllers, while maintaining good impedance tracking performance. The proposed instantaneous model impedance control re-initializes the impedance model to the current position of the robot so the model does not accumulate position errors. The novelty of the method is in using the position feedback both in the outer loop, to track the desired impedance, and in the inner loop to improve robustness. The proposed method includes the dynamic-based and the position-based methods as specific cases, and can be tuned to trade between their corresponding merits. Using simulations, we demonstrate the performance of the new method in comparison with the previous methods. We consider two scenarios: stiff environment, which challenges the position based approach, and load uncertainties, which challenges the dynamic based approach.

Understanding the inherent structure of manipulative hand movements is a critical step toward automatic recognition of manipulative hand movements and toward automatic planning of natural movements for dexterous robots. Manipulative hand movements involve coordinated movements of the digits to manipulate an object within the hand, and are classified as either simultaneous or sequential. Simultaneous hand movements involve a single pattern of coordination, while sequential hand movements involve a sequence of such patterns. Assuming that each coordinated pattern reflects a stable limit cycle of an internal dynamical system, we have previously suggested that stable patterns of coordination are mostly restricted to in-phase or anti-phase coordination. Formulating the coordination underlying simultaneous hand movements reveals that under these conditions the joint-space trajectories are straight lines. We demonstrate the linear nature of the joint-space trajectories projected by the nine basic classes of simultaneous manipulations. The results support our working assumption that in-phase or anti-phase patterns of coordination are prevailing. We conclude that the direction of the trajectory line in the joint space captures the coordination structure underlying a simultaneous hand movement. The vector of direction numbers is extracted and its discriminative power is demonstrated.

Understanding manipulative hand movements is a critical step toward automatic programming of dexterous robot hands, which, given their large number of degrees of freedom, are difficult to program. Our working hypothesis is that the units of manipulative hand movements are coordinated patterns of movement, and that most of the stable patterns of coordination are either in-phase (zero relative phase) or anti-phase (half-a-cycle relative phase). The phase-plane trajectories projected by these patterns of coordination are straight lines, with either positive or negative slope, respectively. In this paper we develop and demonstrate a method for segmenting the phase-plane trajectories, and thereby the movement, accordingly.

The distance between 'texture primitives' is of major interest in characterizing texture images. This is especially natural when the texture primitives are elongated structures aligned in parallel to a common main axis, and the distance is measured along the perpendicular axis. Such images arise, for example, in flow visualization studies, where the elongated structures are low-speed streaks. A point process based texture generation model is developed for the one-dimensional texture along lines perpendicular to the streaks. The point process models the location of the edges of the streaks, and using edge detection techniques, its probability density function can be estimated by the histogram of the distances between the edges. It is shown that for the studied images the resulting histogram is wide (coefficient of variation larger than half), and demonstrated that in this case, previously suggested autocorrelation based methods are not adequate.

No abstract available.