Model parameter learning method

Patent metadata

Field	Value
Publication number	US-12403593-B2
Application number	US-202118042844-A
Country	US
Kind code	B2
Filing date	Apr 30, 2021
Priority date	Aug 27, 2020
Publication date	Sep 2, 2025
Grant date	Sep 2, 2025

Abstract

Provided is a model parameter learning method by which a model parameter of a learning model used in control of a moving body having movement constraints can be appropriately learned. In this model parameter learning method, a model prediction control algorithm reflecting movement constraints of a robot 1 is used to calculate a time series of learning speed commands such that the movement trajectory of the robot 1 tracks the time series of a movement trajectory of a first pedestrian 5 ; and a model parameter of a CNN model is learned by an error back propagation method, the CNN model using learning data including the learning speed commands time series as input and outputting a time series of speed commands for a first moving body.

First claim

The invention claimed is:

1. A robot movement trajectory control method comprising:
a model parameter learning step of learning a model parameter of a learning model and outputting a time series of speed commands of a robot; and
a movement trajectory control step of controlling a movement trajectory of the robot using the learning model and the time series of speed commands of the robot, wherein
the model parameter learning step causes a learning device to learn the model parameter of the learning model used for control of the movement trajectory of the robot such that the robot having a movement constraint moves on the movement trajectory obtained by imitating a movement trajectory of a pedestrian having a movement constraint different from the robot, the model parameter learning step comprising:
an acquisition step of acquiring a time series of surrounding environment information of the pedestrian and a time series of the moving trajectory of the pedestrian synchronized with the time series of the surrounding environment information of the pedestrian;
a calculation step of calculating a time series of learning speed commands of the robot using a model prediction control algorithm based on a model reflecting a movement constraint of the robot such that a time series of the movement trajectory of the robot tracks a time series of the movement trajectory of the pedestrian acquired in the acquisition step; and
a learning step of learning, by a predetermined machine learning algorithm, the model parameter of the learning model, the learning model having data including the time series of learning speed commands of the robot and the time series of the surrounding environment information of the pedestrian input and the time series of speed commands of the robot output.

2. The robot movement trajectory control method according to claim 1 ,
wherein, in the calculation step, a time series of the learning speed commands of the robot is calculated as discrete values.

3. The robot movement trajectory control method according to claim 2 ,
wherein, in the acquisition step, the movement trajectory of the pedestrian that is acquired in a state in which the movement trajectory of the pedestrian does not satisfy the movement constraint of the robot is deleted.

4. The robot movement trajectory control method according to claim 1 ,
wherein, in the acquisition step, the movement trajectory of the pedestrian that is acquired in a state in which the movement trajectory of the pedestrian does not satisfy the movement constraint of the robot is deleted.

Assignees

• Honda Motor Co Ltd

Inventors

• Matsuzaki Sango
• AONUMA Shinta
• Hasegawa Yuji

Classifications

• G06N20/00

  Machine learning · CPC title

• G05D1/65

  Following a desired speed profile · CPC title

• G05D2101/15

  using machine learning, e.g. neural networks · CPC title

• G05B2219/36184

  Record actions of human expert, teach by showing · CPC title

• G05B2219/40116

  Learn by operator observation, symbiosis, show, watch · CPC title