FOESO-Net: A specific neural network for fast sensorless robot manipulator torque estimation.

Area of Science:

• Robotics
• Control Systems
• Mechatronics

Background:

• Contact torque sensing is crucial for safe human-robot collaboration and collision detection.
• Existing sensorless torque estimation methods struggle with noise and limited bandwidth, hindering fast, nonlinear torque observation.
• Traditional approaches often require complex calculations like joint acceleration and inverse inertia matrices.

Purpose of the Study:

• To develop an advanced sensorless torque estimation method for robot manipulators.
• To address the limitations of linear observers in handling noisy and fast time-varying nonlinear torques.
• To enhance the safety and responsiveness of collaborative robots.

Main Methods:

• A novel fractional-order extended state observer (FOESO) is proposed, utilizing momentum as the benchmark state for estimation.
• A customized network, FOESO-Net, is designed, incorporating fractional-order and dynamic weight-based adaptive learning.
• Parallel neural network architectures are employed for adaptive parameter learning of the FOESO.

Main Results:

• FOESO-Net demonstrates convergent error, as formally analyzed and proven.
• The method effectively estimates contact torque, avoiding complex calculations like joint acceleration.
• Simulations and experiments on a collaborative robot platform validate the proposed approach.

Conclusions:

• FOESO-Net significantly reduces estimation error and response time compared to existing methods.
• The proposed fractional-order observer offers superior performance for nonlinear, fast time-varying torque observation.
• This advancement improves the safety and efficiency of human-robot interaction.