Task-Agnostic Exoskeleton Control Using Joint Moment Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current state-of-the-art exoskeleton and motorized prosthesis control systems are task-specific, requiring expensive and limited datasets, leading to limited adaptability and usability for a small set of tasks, and lack the ability to accommodate diverse human behaviors.
Innovation Solution
A task-agnostic exoskeleton control system utilizing deep neural networks for instantaneous biological joint moment estimation, employing a deep domain adaptation method to translate simulated sensor data to real-time control, and a transfer learning framework to optimize user-independent control across various tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If task-specific datasets are used for control system optimization, then control accuracy for specific tasks is improved, but adaptability to diverse human behaviors deteriorates
Solution Approach 1:
The patent applies universality by training the neural network controller on a diverse dataset encompassing multiple tasks and human behaviors rather than task-specific data. The controller learns universal control policies that generalize across different activities, achieving both accuracy and adaptability. The system processes various sensor inputs (IMU, encoder, pressure) to provide context-independent control assistance.
Solution Approach 2:
The patent changes the parameter of training data diversity by using extensive able-bodied dataset with simulated sensors equivalent to the prosthesis domain. This parameter change enables the controller to learn robust representations that generalize across tasks and users without requiring task-specific optimization.
2Reliability
If extensive experimentation is performed for dataset compatibility optimization, then control performance for specific tasks is improved, but development time and cost increase
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network controller on extensive simulated data before deployment. The transfer learning framework pre-processes the learning task by adapting the controller to the specific prosthesis domain using simulated sensor data, eliminating the need for extensive experimentation during actual deployment phases.
Solution Approach 2:
The patent uses copying by creating simulated sensor data that replicates real-world prosthesis operation conditions. The simulation environment copies the essential dynamics and sensor characteristics, allowing the controller to be trained and optimized in silico before real-world deployment, significantly reducing development time.
3Measurement precision
If task-specific controllers are deployed, then control accuracy for trained tasks is improved, but usability for untrained tasks deteriorates
Solution Approach 1:
The patent implements universality by designing a single neural network controller that handles multiple tasks and user types. The controller processes diverse sensor inputs and generates appropriate control commands for various activities without requiring task-specific configuration, making the system easy to operate across different scenarios.
Solution Approach 2:
The patent applies dynamics by using a neural network-based adaptive controller that dynamically adjusts its behavior based on real-time sensor inputs. The controller continuously learns and adapts to different tasks and users during operation, providing accurate control for both trained and untrained tasks without manual reconfiguration.
4Adaptability or versatility
If simulated sensor data is translated to real domain using transfer learning, then adaptability to real devices is improved, but data translation complexity increases
Solution Approach 1:
The patent uses an intermediary approach by introducing a domain adaptation layer that translates simulated sensor data to real device domains. This intermediary translation mechanism bridges the gap between simulation and reality, enabling the controller trained on simulated data to effectively operate on real prosthetic devices with different sensor characteristics.
Data Source
AI summary
Exemplary task-agnostic exoskeleton control system and method are disclosed that are task-agnostic utilizing instantaneous estimates of biological joint moments from deep neural networks to assist the user movements. The exemplary control system employs multiple body sections and joints in-the-loop estimation to provide multi-joint assistance operation, e.g., within an autonomous, clothing-integrated exoskeleton. The exemplary control system may deploy a deep domain adaptation (DDA) method configured to translate human movement data between a simulated sensor domain and a real sensor domain.