Multi-agent control system and method

• US 10,409,230 B2
• Filed: 11/01/2016
• Issued: 09/10/2019
• Est. Priority Date: 11/01/2016
• Status: Active Grant

• Current Assignee: Mitsubishi Electric Research Laboratories, Inc. (Mitsubishi Electric Corporation)
• Original Assignee: Mitsubishi Electric Research Laboratories, Inc. (Mitsubishi Electric Corporation)
• Inventors: Grover, Piyush, Elamvazhuthi, Karthik
• Primary Examiner(s): Saavedra, Emilio J
• Application Number:
  US15/340,015
• Publication Number:
  US 20180120775A1
• Time in Patent Office: Days
• Field of Search: 700 33
• CPC Class Codes: B25J 9/1682 Dual arm manipulator; Coordination of several manipulators
• CPC Class Codes: G05B 13/021 in which a variable is automatically adjusted to optimise the performance
• CPC Class Codes: G05B 2219/32015 Optimize, process management, optimize production line
• CPC Class Codes: G05B 2219/39155 Motion skill, relate sensor data to certain situation and motion
• CPC Class Codes: G05B 2219/39163 Formation control, robots form a rigid formation, fixed relationship
• CPC Class Codes: G05D 1/0027 involving a plurality of vehicles, e.g. fleet or convoy travelling fleet control of land vehicles from a control room G05D1/0297; traffic control systems for road vehicles G08G1/00; for marine craft G08G3/00; for aircraft G08G5/00

Abstract

Motion of multiple agents with identical non-linear dynamics is controlled to change density of the agents from the initial to the final density. A first control problem is formulated for optimizing a control cost of changing density of the agents from the initial density to the final density subject to dynamics of the agents in a density space. The first control problem, which is a non-linear non-convex problem over a multi-agent control and a density of the agents, is transformed into a second control problem over the density of the agents and a product of the multi-agent control and the density of the agents. The second control problem is a non-linear convex problem that is solved to produce the control input for each section of the state space. A motion of each agent is controlled according to a control input corresponding to its section of the state space.

Claims

1. A system for controlling a motion of multiple agents with identical non-linear dynamics, wherein an agent is a device capable to move in a controlled manner, comprising:

a receiver to receive data indicative of at least one of an initial density of the agents in a state space and a final density of the agents in the state space at a given time, or a combination thereof, wherein the state space is partitioned into a set of sections, such that a density of a section of the state space is proportional to a number of agents having states within the section of the state space;

a processor configured to formulate a first control problem optimizing a control cost of changing density of the agents from the initial density to the final density subject to dynamics of the agents in a density space, wherein the first control problem is a non-linear non-convex control problem over a multi-agent control and a density of the agents, wherein the multi-agent control is a function parameterized on the set of sections of the state space defining a control input for each section of the state space;

transform the first control problem into a second control problem, wherein the second control problem is a non-linear convex problem over the density of the agents and a product of the multi-agent control and the density of the agents;

and solve the second control problem to produce the control input for each section of the state space;

and

a controller to control a motion of each agent according to the produced a control input corresponding to its section of the state space.

2. The system of claim 1, wherein the processor is configured to represent the first and the second control problems as a sequence of flows of the agents between neighboring sections of the state space;

solve the second control problem to produce time-varying flow rates between the sections of the state space and time-varying densities of the sections of the state space; and

determine the control inputs at each control step from the time-varying flow rates between the sections and the time-varying densities of the sections.

3. The system of claim 2, wherein the processor is further configured to extract control amplitudes for each section at each time step of the control from the time-varying flow rates between the sections of the state space using the time-varying densities of the sections of the state space;

determine a net control amplitude for each section of the state space in each feasible control direction at each time step of control by combining corresponding control amplitudes; and

determine the control input in each feasible control direction at each time step of control as the corresponding net control amplitude.

4. The system of claim 2, wherein the processor is configured to represent the first and the second control problems by determining a graph structure of the sections of the state space, wherein each vertex of the graph represents a section and neighboring vertices are connected by at least one edge;

and representing the sequence of flows of the agents between neighboring sections of the state space at each time of control as a result of an advection of the densities of the agents on the graph.

5. The system of claim 1,

wherein the first control problem switches between uncontrolled dynamics of the agents and discrete time controls of the agents, wherein the uncontrolled dynamics of the agents are linear in the density space, wherein the discrete time controls of the agents are non-linear in the density space, wherein the uncontrolled dynamics of the agents include a product of a linear transfer operator of the non-linear dynamics of the agents with the density of the agents, and wherein the discrete time controls of the agents include a product of the multi-agent control with the density of the agents;

and

wherein the second control problem replaces the product of the multi-agent control and the density of the agents with a variable to form the non-linear convex problem.

6. The system of claim 5, wherein the processor solves the second control problem for constant frequency of switching times.

7. The system of claim 5, wherein the discrete time controls of the agents are at least ten times faster than the uncontrolled dynamics of the agents.

8. The system of claim 5, wherein the processor is configured to determine a graph structure of the sections of the state space, wherein each vertex of the graph represents a section and neighboring vertices are connected by at least one edge;

and represent the discrete time control of the agents for each section at each time of control as a result of an advection of the densities of the agents on the graph.

9. The method of claim 1, wherein the agents follow non-linear control-affine dynamics,

wherein the first control problem includes a combination of a linear first term and a non-linear second term, wherein the first term includes a product of a linear infinitesimal generator of non-linear uncontrolled dynamics of the agents with the current density of the agents, and wherein the second term includes a product of the multi-agent control, a linear infinitesimal generator of control vector fields, and the density of the agents;

and

wherein the second control problem replaces the product of the multi-agent control for each control vector field and the density of the agents with a variable to form the non-linear convex problem.

10. The system of claim 9, wherein the processor is configured to determine a graph structure of the sections of the state space, wherein each vertex of the graph represents a section and neighboring vertices are connected by at least one edge, wherein each edge is associated with a weight vector having weights given by the linear infinitesimal generator of control vector field;

and represent the control of the agents for each section at each time of the control as a result of an advection of the densities of the agents on the graph.

11. The system of claim 1,

wherein the first control problem includes a product of convex linear combination of different control transfer operators with the density of the agent, wherein the control transfer operators are a result of discretizing allowable controls into a finite number of values;

and

wherein the second control problem replaces the product the density of the agents and weights of the convex linear combination with a variable to form the non-linear convex problem.

12. The system of claim 1, wherein the controller comprises:

a transmitter to transmit a command to each agent to apply a control input corresponding to its section of the state space.

13. The system of claim 1, wherein the controller comprises:

an actuator to apply a control input corresponding to its section of the state space.

14. The system of claim 13, wherein the actuator includes a magnetic actuator for applying a magnetic field to the agents, wherein the magnetic field is determined by the control inputs.

15. The system of claim 1, wherein the receiver receives the initial states of the agents, and wherein the processor determines the initial probability density of the agents in the state space from the initial states of the agents.

16. A method for controlling a motion of multiple agents with identical non-linear dynamics, wherein an agent is a device capable to move in a controlled manner, comprising:

receiving data indicative of at least one of an initial density of the agents in a state space and a final density of the agents in the state space at a given time, or a combination thereof, wherein the state space is partitioned into a set of sections, such that a density of a section of the state space is proportional to a number of agents having states within the section of the state space;

formulating a first control problem optimizing a control cost of changing density of the agents from the initial density to the final density subject to dynamics of the agents in a density space, wherein the first control problem is a non-linear non-convex control problem over a multi-agent control and a density of the agents, wherein the multi-agent control is a function parameterized on the set of sections of the state space defining a control input for each section of the state space;

transforming the first control problem into a second control problem, wherein the second control problem is a non-linear convex problem over the density of the agents and a product of the multi-agent control and the density of the agents;

solving the second control problem to produce the control input for each section of the state space;

and controlling a motion of each agent according to the produced control input corresponding to its section of the state space, wherein at least some steps of the method are performed using a processor.

17. The method of claim 16, further comprising:

representing the first and the second control problems as a sequence of flows of the agents between neighboring sections of the state space;

solving the second control problem to produce time-varying flow rates between the sections of the state space and time-varying densities of the sections of the state space; and

determining the control inputs at each control step from the time-varying flow rates between the sections and the time-varying densities of the sections.

18. The method of claim 17, further comprising:

extracting control amplitudes for each section at each time step of the control from the time-varying flow rates between the sections of the state space using the time-varying densities of the sections of the state space;

determining a net control amplitude for each section of the state space in each feasible control direction at each time step of control by combining corresponding control amplitudes; and

determining the control input in each feasible control direction at each time step of control as the corresponding net control amplitude.

19. The method of claim 17, further comprising:

determining a graph structure of the sections of the state space, wherein each vertex of the graph represents a section and neighboring vertices are connected by at least one edge; and

representing the sequence of flows of the agents between neighboring sections of the state space at each time of control as a result of an advection of the densities of the agents on the graph.

20. A non-transitory computer readable storage medium embodied thereon a program executable by a processor for performing a method, the method comprising:

receiving data indicative of at least one of an initial density of the agents in a state space and a final density of the agents in the state space at a given time, or a combination thereof, wherein an agent is a device capable to move in a controlled manner, wherein the state space is partitioned into a set of sections, such that a density of a section of the state space is proportional to a number of agents having states within the section of the state space;

formulating a first control problem optimizing a control cost of changing density of the agents from the initial density to the final density subject to dynamics of the agents in a density space, wherein the first control problem is a non-linear non-convex control problem over a multi-agent control and a density of the agents, wherein the multi-agent control is a function parameterized on the set of sections of the state space defining a control input for each section of the state space;

transforming the first control problem into a second control problem, wherein the second control problem is a non-linear convex problem over the density of the agents and a product of the multi-agent control and the density of the agents;

solving the second control problem to produce the control input for each section of the state space;

and controlling a motion of each agent according to the produced control input corresponding to its section of the state space.