Laboratory for Perceptual Robotics
Research Archives



The Laboratory for Perceptual Robotics has conducted a range of research focused on technologies for controlling adaptable robot systems that function in open environments.

A Basis for Robot Control, Learning, and Skill Acquisition

This research explores the potential for sensory and motor behavior to be expressed as sequences of control drawn from a control basis . Our approach is related to a body of work aimed toward exploiting `local' and multiple-model approaches to coping with strongly nonlinear and time-varying systems. The control basis structures behavior and provides a formal basis for scheduling resources in adaptive systems. Control actions consist of combinations of closed-loop control primitives with specific resource designations. By enumerating controllers that engage a variety of resources, we support sensorimotor policies that are robust over a wide range of contexts. Moreover, this framework brings traditions in real-time programming and operations research to bear on adaptive electromechanical systems. Solutions to sensorimotor problems are sequences of concurrent control processes designed to traverse a finite set of system equilibria. End-to-end behavior can thus be modeled as a discrete event system, permitting the application of powerful formal techniques to further structure the behavior of the system. Reversible constraints on control actions are used to define legal composite behavior from which a programmer may write behavioral programs. Techniques for automatically programming behavior may also exploit these forms of structure.

Kinodynamic Planning and Control

Connolly, Grupen

An artificial Hamiltonian is constructed using the kinetic energy in the manipulator and the artificial harmonic potential to regulate the overall system energy. Level sets in the energy field define regions to which the system is confined. The energy-referenced controller can be used to plan dynamically feasible paths that reach potential energy states at the reference total energy only with non-zero velocity.

Publications: Connolly, Grupen, and Souccar (1995)

An interactive demo of this approach can be found here .

Motion Planning Accelerators

Connolly, Burleson, Grupen

In conjunction with Prof. Burleson in the ECE department at UMass, Amherst, we are investigating the potential for exploiting parallelism in motion planning algorithms. Estimates indicate considerable promise for real-time path planning. With improved bandwidth, the path planner is capable of tracking moving constraints. automatically incorporating new constraints, and continually updating the underlying harmonic potential surface.

Publications: Grupen et al. (1995) ,

Interpreting Kinesthetic Information

Huber, Grupen

This work is related to interpreting tactile and kinesthetic information during manipulation tasks. Rather than tactile arrays, we are studying the sensory signature of a contact in position, velocity, and force domain information. Contact hypotheses can be derived by recognizing displacements of the manipulator which are consistent with hypothetical geometric constraints. These hypotheses can be corroborated by identifying loads in the manipulator which are consistent with reaction forces at the contact hypotheses. The algorithm we have developed yields estimates for contact location, normal velocity (rolling and sliding) and contact force. This information is valuable in grasp control as well as force control in whole arm manipulators, and can serve as the basis for object recognition strategies.

Publications: Huber and Grupen (1994)

An interactive demo of this approach can be found here .

Haptic Context-Recovery for Grasping and Manipulation

Coelho, Grupen

Robust control techniques are widely applicable to systems in which plant parameters cannot be identified or vary, or whose state cannot be assessed precisely. Under these circumstances, it is desirable to design a controller that is robust with respect to a range of possible plants. Normally, the robust control community considers variations in the dynamics of the plant, we have treated the design of a grasp "controller" that is robust with respect to variations in the geometry of the object. The grasp controller is realized as an adaptive, closed-loop transformation from contact position and normal to stable grasp configurations. We show that equilibria in the composite controller correspond to optimal contact configurations for 2 and 3 contacts on regular, convex polygons and how more general geometries can be addressed. The haptic context of a grasp is determined by matching the dynamic behavior of the grasp controller to pre-established categories. This information reveals the geometric class of the object and can be used to address control compensation, object recognition, and contact allocation.

Publications: Coelho and Grupen (1997) , Coelho and Grupen (1996) , Grupen et al. (1995) , Coelho and Grupen (1994)

MPEG demonstrations of this work can be found here

Legged Locomotion

MacDonald, Huber, Grupen

This work investigates a distributed control approach to legged locomotion that constructs behavior on-line by activating combinations of reusable feedback control laws drawn from a control basis. Sequences of such controller activations result in flexible aperiodic step sequences based on local sensory information. Different tasks are achieved by varying the composition functions over the same basis controllers, rather than by geometric planning of leg placements or the design of new task-specific behaviors. In addition, the device-independent nature of the control basis allows its generalization not only over task domains, but also over different hardware platforms. To show the applicability of this approach, a control basis and two generic control gaits for four-legged walking are introduced and tested on an even terrain walking task in an unknown environment.

Publications: Huber, MacDonald, and Grupen (1996)

Thing, a four-legged walker, is the experimental platform in which this work was implemented.

MPEG demonstrations of this work can be found here

Assembly and Manufacturing

Visual Servoing for Assembly

Schnackertz, Ravela, Grupen

A framework for visually-guided assembly operations is advanced in which a finite state supervisor manages the interaction between control procedures and visual tracking. Unary and binary image plane constraints are used to specify visual servoing tasks. These controllers map image plane feature errors to actuator commands through visuomotor Jacobians. A binocular vision system was constructed to combine image-based feature tracking with weak camera-object calibration techniques to execute peg-in-hole insertion tasks.

Publications: Ravela et al. (1995), Schnackertz and Grupen (1995)

Scheduling Jobs in Flexible Manufacturing Workcells

Araujo, Huber, Dakin, Grupen

This approach to hierarchical scheduling for assembly in a fully-automated FMS is based on a three-level production system. The system consists of a master production plan, which incorporates new customer requests and makes long term production guarantees, a production flow level, planning and monitoring the flow of subassemblies among workstations, and a scheduler level, which is concerned with the sequencing and dispatching of assembly operations. An independent scheduler at each workstation coordinates robot motions and conveyor transferals to interleave the operations related to different subassembly products. The higher priorities of tardy subassemblies are weighed against immediate opportunities to exploit the current configuration of the workstation and maximize overall throughput. We focus mainly on the scheduler level, where heuristics and combinatorial search are combined to achieve near-optimal exploitation of each workstation's capacity.

More information can be found here

Publications: Araujo et al. (1995)

Uncertainty in Assembly

Dakin, Popplestone

The LPR is developing techniques for generating a fine-motion strategies from assembly plans derived from kinematic constraints alone. Insertion clearances and size tolerances are introduced into the assembly part models in parallel with the manual synthesis of a ``perturbed nominal trajectory'' in contact space.

Fine-motion plans for assembly operations are constructed to be robust with respect to position, velocity, and model uncertainty. A nominal assembly motion plan, computed without regard to the effects of uncertainty, is provided a priori. The ``critical points'' (where collisions and jamming are likely to occur) are identified in the nominal trajectory. The configuration space of the assembly is then linearized about each critical point and a heuristic search is performed to select an appropriate fine-motion strategy. A hybrid control strategy is then generated for traversing each contact state in the sequence.

Publications: Dakin (1994) , Dakin and Popplestone (1993) , Dakin and Popplestone (1992)

Learning Assembly Admittance

Gullapalli, Barto, Grupen

This work demonstrated a direct reinforcement learning approach to peg-in-hole insertions using a Zebra Zero robot. This robot is equipped with a wrist force sensor in addition to position encoders. The goal is to improve the ability of a robot controller to insert the peg into the hole over time by incrementally modifying the controller on the basis of position and force feedback.

An artificial neural network was used to generate an appropriate admittance control model for accomplishing the insertion. Nominal position errors yield motion strategies that are subsequently augmented to compensate for sensed contact forces. The insertion policy was ``shaped'' by beginning with a cylindrical peg and relatively large insertion clearances, and subsequently decreasing insertion clearance and varying the peg and hole geometry. This method performed insertions of square pegs into square holes with 0.002" clearances using a robot with +/- 5mm endpoint precision.

Publications: Gullapalli, Grupen, and Barto (1992)

Assembly Planning

Liu, Popplestone

Researchers in the LPR have advanced an approach to geometric reasoning based on the symmetry groups of geometric features. This representation establishes necessary conditions for spatial relationships to hold, sufficient conditions are derived from detailed geometric form and bounding information. This initial work has led to an extension of the approach addressing the development of compliant control strategies based on the symmetry properties of features to be mated.

The geometry and the topology of the assembly components enforce some hard constraints on the configuration(s) of assembly and the order of assembly. To mitigate the combinatorics of searching for a feasible correspondence between features of different bodies, boundary models are matched against a library of standard compound features. A constraint satisfaction network is established and the feasibility of mating compound features is analyzed using the symmetry groups of features.

Publications: Liu and Popplestone (in press) , Popplestone, Liu, and Weiss (1990)

AutoWeb - graphical process design

Popplestone

AutoWeb is a tool which supports engineering design by coupling the evolution of function to the evolution of form from the conceptual level towards the detailed level of design. An engineered system is treated within AutoWeb as a collection of devices, each with terminals. Interaction between nodes is specified by making connections between terminals. Overall, the devices and their connections form a graph or "web." AutoWeb supports the determination of how the behaviour of the whole system specified in the web depends upon the behaviour of the parts.

AutoWeb is intended to be broader in scope than other systems, in that it is designed to support a range of analysis modes as well as to support code generation to realize the computational component of a particular system. Analysis is performed using external engines which are programs specialized to provide a particular capability. For example, the Mathematica system is a mathematical engine which provides the ability to perform symbolic algebra, calculus and function plotting. The ACIS system is a shape engine which provides the ability to manipulate and display shapes of the kind that are manufactured. These two are coupled by the use of a homogeneous internal representation of expressions which can denote both functional entities (e.g. forces, torques, velocities) and form entities (shapes).