TL;DR
This paper introduces a novel Parallel-Hierarchical Neural Network Model (PHNM) that enables simultaneous learning of both arm trajectory and force control. By integrating Hogan's impedance control with a feedback-error-learning scheme, the PHNM effectively addresses the ill-posed problem of inverse dynamics in redundant musculo-skeletal systems, paving the way for more biologically plausible and adaptable robotic control.
What Problem Does This Paper Address?
Controlling a robotic arm to interact with its environment requires precise coordination of both its position (trajectory) and the forces it exerts. While previous models, like the hierarchical neural network proposed by Kawato et al. (1987), excelled at trajectory control, they struggled to directly adapt to the unique mechanics of human-like musculo-skeletal systems. Hogan's impedance control (1984, 1985) offered a way to control both trajectory and force by specifying virtual trajectories and mechanical impedance, but it didn't inherently explain how this control is learned. A key challenge arises from the redundancy in biological arms: multiple muscles (agonists and antagonists) contribute to joint torque, making it difficult to uniquely determine the muscle commands needed for a desired torque and trajectory. This "ill-posed problem" means that a direct inverse model is not straightforward.
Methods
The core of this paper is the Parallel-Hierarchical Neural Network Model (PHNM). It builds upon previous work by integrating:
- Hogan's Impedance Control: This provides a framework for controlling both trajectory and force by defining desired mechanical impedance and virtual trajectories.
- Feedback-Error-Learning Scheme: This learning mechanism, previously proposed by Kawato et al. (1987), allows the neural network to adjust its internal parameters based on the errors between desired and actual outcomes.
The PHNM features two parallel inverse models:
- Inverse Statics Model (ISM): Primarily for controlling equilibrium posture and mechanical stiffness.
- Inverse Dynamics Model (IDM): For compensating dynamic properties during fast movements.
These models are trained using a feedback-error-learning algorithm where synaptic weights are updated based on the error signal from a feedback controller. The paper also introduces two novel hierarchical control laws derived using the Moore-Penrose pseudo-inverse matrix. These laws address the redundancy issue by optimizing for:
- Minimum Muscle-Tension-Change Trajectory: Minimizing the rate of change of muscle tension.
- Minimum Motor-Command-Change Trajectory: Minimizing the rate of change of motor commands.
These control laws, combined with the feedback-error-learning, resolve the ill-posed problem, allowing the PHNM to acquire the necessary inverse models.
Datasets
Not discussed in the paper. The experiments utilize an artificial muscle arm and computer simulations.
Key Insights and Ideas
- Resolving the Ill-Posed Problem: The combination of hierarchical control laws (optimizing for minimum tension or command change) and feedback-error-learning is crucial for solving the ill-posed inverse dynamics problem in redundant musculo-skeletal systems.
- Biologically Plausible Learning: The PHNM offers a more biologically plausible explanation for motor learning by integrating concepts from impedance control and neural network learning.
- Simultaneous Trajectory and Force Control: The model demonstrates the ability to learn and execute control for both position and force simultaneously, a critical capability for dexterous interaction.
- Hierarchical Control for Redundancy: The two distinct hierarchical control laws highlight how the nervous system might manage redundancy by prioritizing different objectives (minimizing tension change vs. command change) depending on the task (slow vs. fast movement).
Key Results
- Accurate Trajectory Control: The PHNM, particularly with the minimum muscle-tension-change criterion, achieved accurate and smooth trajectory control in slow movements using the artificial muscle arm, outperforming conventional feedback control.
- Stable Fast Movement Control: For fast movements, the minimum motor-command-change criterion enabled stable trajectory control in computer simulations, explaining the observed phenomenon of increased mechanical impedance through coactivation.
- Accurate Force Control: The minimum motor-command-change criterion also proved effective for accurate force control when combined with the trained ISM and IDM.
- Generalization Capability: The trained parallel inverse models showed good generalization capabilities even with untrained trajectories.
- Learning Efficiency: The learning of the ISM and IDM was shown to be completed within approximately 2,000 iterations.
Limitations and Future Work
- Artificial Muscle Arm Limitations: While the artificial muscle arm mimics human arm mechanics, it is still a simplified model. The paper notes that feedforward control using trained ISM and IDM is necessary, implying that purely feedback-based control is insufficient for complex tasks.
- Future Work: The paper implicitly suggests that further exploration of the interplay between different hierarchical control laws and their impact on various movement types would be beneficial. The successful application to an artificial arm also opens doors for testing and refinement on more complex robotic systems and potentially in human-robot interaction scenarios.
Summary updated