Robot Navigation Training With Mixed Exploration on Edge Hardware
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high computational complexity and memory requirements of Reinforcement Learning combined with Deep Neural Networks hinder deployment on low-power, resource-constrained robotic systems for autonomous navigation in unmapped environments.
Innovation Solution
A system and method for training a robot using a neuromorphic processor to record and analyze movement data, employing a mixed exploration strategy that combines random and novelty-based exploration to navigate unmapped environments, minimizing memory footprint and enabling real-time navigation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If Reinforcement Learning combined with Deep Neural Networks is used for autonomous navigation, then navigation capability in unmapped environments is improved, but computational complexity and memory requirements increase significantly
Solution Approach 1:
The navigation system is segmented into multiple specialized modules: a movement model for locomotion planning, a collision model for obstacle avoidance, and a goal model for destination guidance. Each module is trained independently on specific tasks and then integrated, reducing the overall computational complexity compared to training a single large neural network while maintaining navigation capability in unmapped environments.
2Reliability
If Reinforcement Learning with backpropagation is used, then learning effectiveness is improved, but memory footprint and energy consumption increase
Solution Approach 1:
The training process is segmented into separate tasks for each navigation module (movement, collision, goal models). Each module is trained independently with targeted datasets, reducing the total computational burden and memory requirements compared to training a monolithic network. This segmentation enables effective learning while lowering energy consumption on resource-constrained robotic systems.
Data Source
AI summary
A system and a method for training and navigating a robot in an unmapped environment is disclosed. The method includes, training a movement model for the robot by recording and analyzing movement data of the robot as the robot moves over a plurality of walks within an environment, wherein the training is based on a mixed exploration strategy that incorporates a novelty-based exploration factor into a random exploration strategy, solving a planning task using the movement model, evaluating the movement model based on a planning error that represents a distance between the goal location and a final position reached by the robot on solving the planning task and navigating an unmapped environment including a source and a destination based on actions output by the trained movement model that successfully solves the planning task during the evaluation.