Search

CN-122008230-A - Mobile terminal neural network real-time resolving method for parallel mechanism bidirectional kinematics

CN122008230ACN 122008230 ACN122008230 ACN 122008230ACN-122008230-A

Abstract

The invention discloses a real-time calculation method of a mobile terminal neural network of parallel mechanism bidirectional kinematics, which relates to the technical field of robot motion control and comprises the steps of determining an effective joint angle value interval, collecting joint angle samples offline and obtaining end positions to form a forward training sample and a reverse training sample, carrying out sine and cosine joint coding on joint angles, carrying out minimum and maximum normalization and parameter preservation on the end positions and the angle codes, constructing a bidirectional neural network of a shared feature extraction base, a forward output head and a reverse output head, training to obtain a forward kinematics model and a reverse kinematics model, converting the forward kinematics model and the reverse kinematics model into an inference model file, loading the inference model file and normalization parameters in a control period, inputting target end positions to generate joint angle instructions, carrying out candidate screening and first-order low-pass filtering, and reading back joint angles to obtain current end positions and differentiating to obtain end speeds. The method reduces storage and maintains continuity.

Inventors

  • XUE FEI
  • LIANG XIUJIE
  • MEI YONGJUN

Assignees

  • 上海了得科技有限公司

Dates

Publication Date
20260512
Application Date
20260325

Claims (10)

  1. 1. The real-time mobile terminal neural network resolving method for the two-way kinematics of the parallel mechanism is applied to a robot control system, and the leg parallel mechanism is driven by two motor controllers and is characterized by comprising, Acquiring joint angle samples offline in an effective value interval of the joint angles and obtaining corresponding tail end positions through simulation or calibration to form a forward training sample and a reverse training sample; The method comprises the steps of carrying out sine and cosine joint coding on joint angles, carrying out minimum maximum normalization on end positions and angle codes, storing normalization parameters, training a bidirectional neural network based on a shared feature extraction base, a forward output head and a reverse output head to obtain a forward kinematic model and a reverse kinematic model, loading the forward kinematic model, the reverse kinematic model and the normalization parameters in a control period, inputting a target end position into the reverse kinematic model to obtain angle codes, restoring the joint angles, sending the angle codes to a motor controller, inputting a readback joint angle into the forward kinematic model to obtain a current end position, and calculating the end speed.
  2. 2. The method for real-time calculation of a neural network at a mobile terminal according to claim 1, wherein: And determining an effective joint angle value interval according to the stroke of the motor and the geometric constraint of the mechanism, performing zero calibration and direction alignment after assembly, converting the encoder angle reading into joint angles, writing a zero offset and direction matrix into a configuration file, and initializing and loading.
  3. 3. The method for real-time calculation of a neural network at a mobile terminal according to claim 1, wherein: and acquiring a joint angle sample by adopting uniform sampling and random supplementary sampling in an effective joint angle value interval, and increasing sampling density in a working space boundary area to form a boundary-encrypted joint angle sample.
  4. 4. The method for real-time calculation of a neural network at a mobile terminal according to claim 3, wherein: The tip position is obtained by simulation or calibration, including calculating the tip position based on a high-precision simulation model to form tag data, and converting the measured tip position to a base coordinate system by fiducial point alignment when the tag data is acquired using a measurement device.
  5. 5. The method for real-time calculation of a neural network at a mobile terminal according to claim 4, wherein: sample expansion is performed on the joint angle samples and the tail end positions, wherein the sample expansion comprises the steps of superposing angle disturbance on the joint angle samples, superposing position disturbance on the tail end positions, and performing effective value interval filtering on the expanded samples.
  6. 6. The method for real-time calculation of a neural network at a mobile terminal according to claim 1, wherein: The joint angles are converted into angle coding vectors by sine and cosine joint coding, minimum and maximum normalization is respectively carried out on the tail end positions and the angle coding vectors, and the dimension-by-dimension minimum value and the dimension-by-dimension maximum value are used as normalization parameters to be stored and multiplexed in an reasoning stage.
  7. 7. The method for real-time calculation of a neural network at a mobile terminal according to claim 6, wherein: the bidirectional neural network comprises a shared feature extraction base, a forward output head and a reverse output head, and parameters of the shared feature extraction base are updated in combination with parameters of the forward output head and the reverse output head in the same training process.
  8. 8. The method for real-time calculation of a neural network at a mobile terminal according to claim 7, wherein: And converting the trained forward kinematics model and reverse kinematics model into a mobile end reasoning model file, comparing the training end output with the reasoning end output by a verification sample after conversion to execute consistency verification, and binding and storing the normalized parameters and the reasoning model file.
  9. 9. The method for real-time calculation of a neural network at a mobile terminal according to claim 8, wherein: and (3) packaging an inference engine in the control system, reading a configuration file, loading a forward kinematics model, a reverse kinematics model and normalization parameters, entering a simulation mode when a pointer of a motor controller is empty, and otherwise, entering a real hardware mode.
  10. 10. The method for real-time calculation of a neural network at a mobile terminal according to claim 9, wherein: And in the control period, candidate screening and first-order low-pass filtering are carried out on the joint angle output by the inverse kinematics model according to the current joint angle, a control instruction is issued, and the read-back joint angle is converted by the forward kinematics model to obtain the current end position and then the end speed is calculated according to the time difference.

Description

Mobile terminal neural network real-time resolving method for parallel mechanism bidirectional kinematics Technical Field The invention relates to the technical field of robot motion control, in particular to a mobile terminal neural network real-time calculation method of parallel mechanism bidirectional kinematics. Background In humanoid robots, quadruped robots and exoskeletons, the legs generally adopt a planar parallel mechanism as a transmission mechanism near the hip/knee, so that the robot has very strong bearing capacity and dynamic response in a limited space. The kinematic mapping of the mechanism is mostly that the motor rotation angle of a multi-rod closed loop is mapped to the movement of the tail end on an X-Y plane, the feasible domain of the kinematic mapping of the mechanism which is strong in nonlinearity is limited by mechanical travel and geometric constraint, and sensitivity surge and singular approaching are more likely to occur near the boundary of a working space. The leg control is carried out by adopting a 1kHz level period, on-line calculation is carried out by matching with a motor current loop, gesture estimation and upper gait planning, and the embedded controller is limited by calculation force, storage, power consumption and the like and cannot bear the worst time delay of high iteration times and large-scale table lookup. The existing control system is to obtain the end position and the target joint instruction usually by adopting an analytic method, a numerical iteration method or a discrete table lookup/interpolation method, wherein the analytic method is to deduce a closed equation according to specific topology and process multi-solution branches, actual assembly tolerance, compliance and zero offset can cause deviation between an analytic model and a real object, the numerical iteration method needs initial values and Jacobian conditions, and is easy to cause non-convergence, jump solution and uncontrollable iteration times when approaching a boundary or singular, so that calculation delay jitter and control margin are shortened, the discrete table lookup method can avoid iteration, but the coverage universe requires high-density sampling and large-capacity storage, interpolation errors of the boundary and a high curvature area can be amplified, and a jump phenomenon is easy to occur due to discontinuous index of the multi-solution area, and table updating and calibration version management are complex. In addition, the slow drift of zero position caused by sensing noise, return clearance and temperature drift can cause unstable switching of 'numerical solution corresponding to the same end target' in adjacent control periods, further amplifies instruction jitter, and if control is not added, torque peak, structural vibration and tracking error accumulation are easily caused, and overcurrent/overtemperature protection or mechanism interference collision and falling risks are caused in serious cases. Therefore, the technical problem faced by the prior art is how to solve the bidirectional motion relation between the joint angle and the tail end position stably, continuously and reproducibly in a strict real-time control period on the premise of meeting the effective travel constraint and the geometric feasibility constraint of the leg parallel mechanism, and avoid control instruction mutation caused by working space boundary, multi-solution area non-convergence, solution branch jump and interpolation amplification error. Disclosure of Invention (One) solving the technical problems Aiming at the defects of the prior art, the invention provides a real-time calculation method of a mobile terminal neural network of parallel mechanism bidirectional kinematics, which comprises the steps of carrying out sine and cosine joint coding on joint angles, carrying out minimum maximum normalization on terminal positions and angle codes, saving parameters, constructing a bidirectional neural network of a shared feature extraction base, a forward output head and a reverse output head, training to obtain a forward kinematics model and a reverse kinematics model, converting the forward kinematics model and the reverse kinematics model into an inference model file, loading the inference model file and normalization parameters in a control period, inputting a target terminal position to generate a joint angle instruction, carrying out candidate screening and first-order low-pass filtering, and sending back the joint angle to obtain the current terminal position and differentiating to obtain the terminal speed. (II) technical scheme In order to achieve the above purpose, the invention is realized by the following technical scheme: a real-time calculation method of a mobile terminal neural network of a parallel mechanism bidirectional kinematics is applied to a robot control system, a leg parallel mechanism is driven by two motor controllers, and comprises, Acquiring joint angle samples offli