人形机器人控制：带有 ZMP 步态调度器和神经逆运动学的混合信号脚步规划 SoC.pdf

1、2025 Asia and South Pacific Design Automation Conference2025 Asia and South Pacific Design Automation ConferenceHumanoid Robot Control:A Mixed-Signal Footstep Planning SoC with ZMP Gait Scheduler and Neural Inverse KinematicsQiankai Cao1,Yiqi Li1,Juin Chuen Oh1,Jie Gu11Northwestern UniversitySlide 0

2、2025 Asia and South Pacific Design Automation Conference2025 Asia and South Pacific Design Automation ConferenceChallenges of Humanoid Robot ControlComputationally heavy for 3D footstep planning on humanoid robotSpecial trajectory control of the robots center of mass(CoM)for balancingHigh computatio

3、n for 10-20 DoF robot joint control compared with wheeled robotSlide 12F-3House worksInverse Kinematics for Motion Control 3Search and RescueFoot CrossoverCoMyzCoMxyz2ZMP Center of Mass Trajectory for Balancing3D Footstep Planning11:RFoot3:RKnee4:RHipRoll6:RElbow7:RShoulderPitch8:RShoulderRollx2:RAn

4、kle5:RHipPitch2025 Asia and South Pacific Design Automation Conference2025 Asia and South Pacific Design Automation ConferenceContribution of This WorkSlide 22F-3 Time-domain Graph Search Engine for 3D Foot planning Mixed-signal Circuit for Efficient ZMP Pattern Generation Neural Network Approximato

5、r for Inverse Kinematics In-situ Demo on Humanoid RobotVTXVTXVTXVTXVTXVTXVTXVTXVTXVTXVTXVTXVTXVTXVTXVTXVTXVTXVTXVTX.Time-domain Neural NetworkDigital Neural Network1x2x4x8x1x2x4x8x1x2x4x8xF(x)F(x)F(x)F(x)F(x)zmp=mg(x+px)-mzc=0 EN.vco_tuneDemo boardDriver boardFront view of robotBack view of robot an

6、d foot stepsLeft footRight footScan cableMini FPGAMotor CAN bus2025 Asia and South Pacific Design Automation ConferenceTime-Domain Computing(TC)Key benefits:energy/area efficiency,digital implementationTime EncoderDinn:0Time LogicOutn:0TDTime DecoderTEA3:0B3:0InC3:0D3:0InIn1In2COMPARECMP(A+B,C+D)Out