Robot Motion Transfer Using Learned Link-Direction Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Robotic systems face challenges in mimicking human motion due to differences in dimensionalities and kinematic structures, leading to difficulties in controlling joints and maintaining balance, especially when transitioning from highly dimensional human movements to lower dimensional robotic movements.
Innovation Solution
The use of a non-linear approximator, such as a neural network, is trained to generate joint control signals by learning from motion capture data, utilizing a reward function that compares the directions of robot links to target links, allowing the robot to mimic human-like motions while preserving balance, even with fewer degrees of freedom.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a robot controller attempts to mimic human motion with high degrees of freedom, then the motion fidelity to human movement improves, but the control complexity and difficulty of preserving balance increase due to dimensional mismatches between human and robot joint structures
Solution Approach 1:
The patent introduces an intermediary mapping system that translates human motion capture data into robot control commands through a dimensionality adaptation layer. This intermediary handles the dimensional mismatch between human and robot joint spaces, allowing high-fidelity motion reproduction without directly exposing the control system to the full complexity of human kinematics.
Solution Approach 2:
The system dynamically adjusts control parameters by transforming motion data from human joint space to robot joint space through learned mapping functions. This parameter transformation allows the robot to reproduce human motions while adapting to its own mechanical constraints and degree of freedom limitations.
2Ease of manufacture
If the robot joint has fewer degrees of freedom compared to the target human joint, then the mechanical and control constraints are reduced, but the ability to accurately replicate human motion is hindered
Solution Approach 1:
The patent resolves the degree of freedom mismatch by introducing additional control dimensions through coordinated motion of multiple robot joints. Instead of requiring one-to-one joint correspondence, the system distributes the motion task across multiple joints, effectively creating virtual degrees of freedom through cooperative control that compensates for the robot's mechanical limitations.
Solution Approach 2:
The mapping process segments the human motion into discrete components that can be independently adapted to robot capabilities. By breaking down complex human joint movements into smaller motion elements, the system can redistribute these elements across available robot joints, maintaining motion accuracy despite fewer degrees of freedom.
3Device complexity
If traditional motion transfer methods are used without reinforcement learning, then the training process is simpler, but the robot's ability to adapt to dimensional constraints and maintain balance deteriorates
Solution Approach 1:
The patent implements feedback loops where the robot's balance state and motion accuracy continuously inform the reinforcement learning agent. This feedback mechanism allows the system to learn optimal control strategies that maintain balance while adapting to dimensional constraints, with the reward function guiding the agent toward solutions that satisfy both motion fidelity and stability requirements.
Solution Approach 2:
The system employs dynamic adaptation through reinforcement learning, allowing the control policy to evolve and adjust in real-time based on environmental feedback. This dynamic approach enables the robot to maintain balance during motion execution by continuously optimizing control commands, rather than relying on static pre-computed trajectories.
Data Source
AI summary
Techniques for transferring highly dimensional movements to lower dimensional robot movements are described. In an example, a reference motion of a target is used to train a non-linear approximator of a robot to learn how to perform the motion. The robot and the target are associated with a robot model and a target model, respectively. Features related to the positions of the robot joints are input to the non-linear approximator. During the training, a robot joint is simulated, which results in movement of this joint and different directions of a robot link connected thereto. The robot link is mapped to a link of the target model. The directions of the robot link are compared to the direction of the target link to learn the best movement of the robot joint. The training is repeated for the different links and for different phases of the reference motion.