Neural Network Motion Control via Surface Damping Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning techniques for motion control in animation and robotics produce unnatural movements and fail to account for real-world constraints, making them unsuitable for controlling physical robots, especially in conditions requiring balance and walking mechanics.
Innovation Solution
Training Neural Networks (NNs) to accept sensor readings and servo control commands as inputs, with integration and backlash threshold comparisons to simulate real-world constraints, using reinforcement learning and various NN architectures like RNNs and CNNs to generate smooth, life-like motions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning techniques are used to generate character movements, then the need for human animators is reduced, but the movements become unnatural and jittery
Solution Approach 1:
The patent implements feedback mechanisms where the neural network's output movements are continuously evaluated against physical constraints and real-world robot dynamics models. The system uses reinforcement learning with reward functions that penalize unnatural movements and encourage physically plausible motion, creating a closed-loop training process that iteratively improves movement naturalness while maintaining automation.
Solution Approach 2:
The patent transforms the neural network's output parameters from direct joint positions to joint velocities or accelerations, and introduces additional parameters representing physical constraints (friction, inertia, contact forces). This parameter transformation allows the network to learn more natural movement patterns that respect physical realities, reducing jitter while preserving the benefits of automated generation.
2Extent of automation
If machine learning animation models are used, then animation creation is automated, but real-world constraints are not accounted for
Solution Approach 1:
The patent incorporates real-world constraints and physical models into the training process before deploying the neural network for实际控制. During the training phase, the system pre-learns the dynamics of the physical robot, including friction characteristics, inertia properties, and contact mechanics, embedding this knowledge into the network's decision-making process so that the automated controller is already adapted to real-world conditions.
Solution Approach 2:
The patent introduces a physics-based dynamics model as an intermediary between the neural network and the physical robot. This intermediary layer translates the network's high-level motion intentions into physically feasible commands while accounting for real-world constraints, effectively mediating between automated decision-making and physical reality without requiring direct trial-and-error on the actual robot.
3Reliability
If hand animation is used to create life-like movements, then movement quality is high, but the process is time consuming and requires artistic skill
Solution Approach 1:
The patent enables the animation system to self-adjust and self-optimize by using reinforcement learning where the neural network automatically learns optimal movement parameters through trial and error in simulation. The system serves itself by autonomously improving movement quality based on performance feedback, eliminating the need for manual artistic adjustment while maintaining high-quality, natural-looking motions.
Solution Approach 2:
The patent uses motion capture data from real human or animal movements as training examples for the neural network. By copying and learning from naturally occurring motion patterns, the system acquires authentic movement characteristics without requiring manual animation expertise, significantly reducing creation time while preserving the quality of natural motion.
4Reliability
If motion capture is used to model character movements, then realistic motion is achieved, but it is expensive and time consuming for multiple characters
Solution Approach 1:
The patent creates a universal neural network model that can be applied to multiple different character types and robot configurations. Instead of creating separate motion models for each character, the system learns general movement principles that can be adapted to various anatomies and mechanical structures, making the solution scalable and cost-effective for controlling many different characters or robots with a single trained model.
Data Source
AI summary
A method, system and computer product for training a control input system involve taking an integral of an output value from a Motion Decision Neural Network for one or more movable joints to generate an integrated output value and generating a subsequent output value using a machine learning algorithm that includes a sensor value and a previous joint position if the integrated output value does not at least meet the threshold. Surface damping interactions with at least a simulated environment, a rigid body position and a position of the one or more movable joints based on an integral of the subsequent output value are simulated. The Motion Decision Neural Network is trained with the machine learning algorithm based upon at least a result of the simulation of the simulated environment and position of the one or more movable joints.