Robot Drive Using HDC Vectors for Energy-Efficient Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning methods for robot systems face inefficiencies in energy consumption and performance due to the inconsistency between energy limitations and calculation demands, necessitating a more efficient and lightweight learning approach.
Innovation Solution
A robot drive system utilizing high-dimensional vector computing (HDC) that encodes time-series sensor data into high-dimensional input vectors, employs imitation learning with hard negative mining, and reinforcement learning to generate robot control signals, improving learning speed and energy efficiency while maintaining high accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If conventional machine learning methods are used for robot control, then calculation accuracy can be maintained, but energy consumption increases and calculation speed decreases
Solution Approach 1:
The patent replaces conventional numerical calculation methods (Boolean logic-based arithmetic operations) with high-dimensional vector computing that uses binding operations. This substitution fundamentally changes the computational paradigm from traditional CPU-based arithmetic to a system that can leverage specialized hardware accelerators, thereby reducing energy consumption while maintaining or improving calculation speed for machine learning inference in robot control
Solution Approach 2:
The patent transforms the representation of sensor data and control signals from conventional numerical values to high-dimensional vectors. By changing the parameter representation from scalar/numerical to high-dimensional vector space, the system enables more efficient binding operations that consume less energy and can be parallelized more effectively, improving both energy efficiency and calculation speed
2Productivity
If the latest GPU is used for machine learning, then calculation performance improves, but energy consumption increases
Solution Approach 1:
The patent substitutes GPU-based floating-point arithmetic operations with high-dimensional vector binding operations that can be performed on specialized hardware. This replacement targets the specific bottleneck of energy-intensive arithmetic operations in neural network inference, achieving comparable or superior calculation performance with significantly reduced energy consumption by using hardware-optimized binding operations instead of general-purpose GPU computation
3Measurement precision
If conventional machine learning methods are used, then learning accuracy can be achieved, but operation efficiency decreases
Solution Approach 1:
The patent replaces conventional training and inference procedures with high-dimensional vector binding operations throughout the machine learning pipeline. By substituting arithmetic-based neural network operations with binding-based vector operations, the system achieves the same learning accuracy while dramatically improving operation efficiency, as binding operations can be executed more quickly and with less computational overhead than traditional arithmetic operations
Data Source
AI summary
A robot drive system includes: an encoder for encoding a plurality of pieces of time-series sensor data measured from a plurality of sensors provided in a robot to generate a plurality of high-dimensional input vectors; an imitation learning unit for training a high-dimensional computing (HDC) object model to generate a high-dimensional vector matched with each of a plurality of robot actions and sensor information using a learning data set including a plurality of pieces of sensor information of the plurality of sensors labeled with a control action signal indicated by a robot drive signal received from a user in response to each point in time; and a similarity analyzer for analyzing a similarity between a plurality of high-dimensional object vectors generated from the HDC object model and the plurality of high-dimensional input vectors when training of the HDC object model is completed.