Sampling-Based Nonlinear Stochastic Optimal Control for Neuromechanical Systems.

Area of Science:

• Neuroscience
• Robotics
• Control Theory

Background:

• The nervous system's control of tendon-driven movements is not fully understood.
• Stochastic Optimal Control (SOC) is a proposed model for neural control, involving solving the Hamilton-Jacobi-Bellman equation to minimize cost functions with noisy inputs.

Purpose of the Study:

• To evaluate and compare three SOC methodologies for controlling a simulated 3-joint human index finger.
• To determine which SOC methods best handle kinematic redundancy and noise in biological systems.

Main Methods:

• Simulated a planar 3-joint human index finger performing a tapping task.
• Implemented and compared three SOC algorithms: iterative Linear Quadratic Gaussian (iLQG), Model-Predictive Path Integral Control (MPPI), and Deep Forward-Backward Stochastic Differential Equations (FBSDE).
• Averaged results over 128 simulation repeats to assess performance and variability.

Main Results:

• All three methods could position the fingertip at desired final joint angles.
• MPPI and FBSDE demonstrated superior performance compared to iLQG, exhibiting less kinematic variance and deviation from target joint angles.
• Kinematic redundancy and noise resulted in different joint trajectories and final postures across the methods.

Conclusions:

• MPPI and FBSDE are more effective than iLQG for controlling nonlinear, tendon-driven systems with inherent noise.
• The mathematical frameworks of MPPI and FBSDE are potentially applicable to tendon-driven robots, exoskeletons, and prosthetic devices.