21
1 生生生生生生生生 生生生 HOAP-1 Biologically Inspired locomotion control & HOAP-1 bipedal walking Autonomous System Lab, Fujitsu Lab

cpg

Embed Size (px)

DESCRIPTION

cpg2

Citation preview

  • Biologically Inspired locomotion control & HOAP-1 bipedal walkingAutonomous System Lab, Fujitsu Lab

  • OutlineTraditional legged robot control & Biological-inspired control strategyNeural mechanism for CPG-based locomotionSensory integration and reflex for spinal controllerForward dynamics based simulationHOAP-1 walkingPrimary results and future works

  • Locomotion in Legged RoboticsControl strategies for legged locomotion

    Model based planBehavior based reaction chainLimit-cycle based Dynamics (CG)Passive walkingProblems?

  • Biological inspired robotics

    Not duplication of the total bio-systems. That would not be possible. The main points are:Take the knowledge from sensory and motor research in biological systems implement that knowledge in the design of our biologically inspired robots. The biologically inspired approach provides for flexible and versatile robotic systems. It also provides a basis for intelligent control and autonomous behavior. Engineering autonomous humanoid robots will help us really understand the nature of human locomotion control So, whats the difference?

  • Cyclic locomotion mechanismNeural systemMotor commands:1 descending input to PG 2 PG motor commands3 motoneuron outputs4 descending supraspinal motor commandsFeedback pathways:5 local spinal reflex pathways6 reflex pathways to PG7 supraspinal reflex pathways8 ascending signals from SSC to PG9 ascending signals from SSC to supraspinal center10 ascending PG outputModulatory pathways11 supraspinal modulation of PG reflexes12 PG modulation of spinal reflexes13 supraspinal modulation of spinal reflexes

  • Pattern Generator (PG)Also called Central Pattern Generator (CPG), is a group of nonlinear oscillatory neurons that is capable of generating oscillatory output in the absence of phasic inputs.

    They are governed by:

  • Control mechanism of CPGconstant rhythmic signal generatorthe regulation of the tonic stimulus intensity of the whole inputs;the temporary change of input in the networks potentially producing more than one rhythm;the change of synaptic weights;the alteration of part of stimuli; The target of using CPG is: get the right timing! Most important and interesting property:Entrain the outside inputs!Sin inputCPG outputSpecial emphasis on some very recent developments which support the view that there is a human spinal CPG for locomotion.

  • Motoneuron & Peripheral feedbackMajor sensory system in locomotionVision systemVestibular systemSomatosensory systemMotoneuron modelWhy need motoneuron model?(Amplitude)CPG shaperGeneral framework for reflex integrationController flexibility enhancementBiological plausible

  • State dependant sensory feedbackThe sensory feedback to spinal controller is state dependant, the different walking state is decided by vestibular sensor and COM-COG relationship.

  • Multibody dynamics systemThe mechanical system has a mind of its own, governed by the physical structure and laws of physics. (Raibert & Hodgins) Contact force, ground static&kinetic fiction, joint fiction, and passive visco-elastic joint limit integratedFeatherstone's method ( O(n) ) for multibody forward dynamics.

  • System constructionRobot leg DOF configuration:Hip 2(pitch, roll), Knee 1 (pitch), Ankle 2(pitch, roll)CPGs: Coupled with each joint, total 10 CPGsMotoneurons and virtual muscles: According to the simplified anatomy muscle model, total 18 motoneurons, each drives one muscle.

  • Forward dynamics optimization

  • What we learned (1) : Total SystemHigh dimensional locomotion control system (humanoid robot) are closely coupled!Motor output is constantly modified by both neural and mechanical feedback.CPG can not work along for such a total system although this idea is conflict with the definition of CPG (is it really Central?)Neural signals are not commands but suggestions sent to a mechanical system possessing its own behavior realized through its physical interaction with the environment.Because neural and mechanical systems are dynamically coupled to each other and to the environment, it is not always clear what is the controller and what is the plan.VisionhearingTouchContactStrainBalance

    Visco-elastic response

  • What we learned (2)Main difficultiesConstruct a total couple dynamical system is difficult, even in the simulation case!1 Nonlinearities in both the neural and robot dynamics.2 Parameter sensitivities make the stable controller tuning difficult.3 Using stochastic method (GA) maybe is not the best way in this case. But there lacks the efficient algorithms for nonlinear programming.4 The CPG&MN model we used are sufficient? The amount of neurons, the topological structure and the coupling format5 Muscles are more than motors, they also serve as brakes, springs, and struts. But we can not grasp all of the properties and take advantages of them.

  • So what should we doSo, instead of trying to get the total system in the very beginning, it is better first to construct a running system and then add features step by step. It will also help to understand the details about the internal neural system components well.

    Bottom-Up strategy to make HOAP-1 walk his first step.

  • What is the bottom?Basic neuron functions:Linear summationConstant multiplyingSwitchDelay

    Also we choose second order differential equations for its prevailing advantages in understanding and description of neural dynamics.

  • New governing Equations

    where:

    and

    ,

    ,

    are predefined constants.

    _1064902143.unknown

    _1064910149.unknown

    _1064910157.unknown

    _1064910124.unknown

    _1064901935.unknown

    _1064909679.unknown

    _1064909982.unknown

    _1064910070.unknown

    _1071900924.unknown

    _1064909812.unknown

    _1064909193.unknown

  • Neural model for HOAP-1The neural model can be cataloged as joint-based or motion primitive-based

  • Sensory implementationSensory feedback input to the neural system through a delay, which is governed by a first order differential equation for natural integration of the total system. Now, the foot force sensors are integrating in the neural system to simulate the function of tactile receptor reflex.

  • Implementation issuesGoverning equations are integrated in real-time;The parameters for different motion primitives are tuned in a straightforward manner.The compensation for gear&motor position error.Walking patterns vary from turning, walk, and turning+walking.

  • Future works Analytic work towards understanding CPGNonlinear dynamics & Nonlinear Programming. New motor primitives equations.

    HOAP-1 walking performance improvementSpecial investigation for sensory system. Different walking environment.Real-time software enhancement.The future works will focus on: