Robot Visual Motion Prediction for Action-Conditioned Planning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for robot interaction with the environment require manually labeled object information, making it impractical to scale real-world interaction learning to various scenes and objects, and existing video prediction techniques struggle with complex and stochastic physical interactions, particularly in real-world scenarios.

Innovation Solution

An action-conditioned motion prediction model that explicitly models pixel motion, predicting a distribution over pixel motion from previous frames, allowing robots to predict object motion without relying on object appearance, using datasets of robot interactions to learn physical interactions autonomously.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manually labeled object information is used for robot interaction learning, then prediction accuracy can be improved, but the complexity and cost of data preparation increases significantly

Engineering Contradiction:
Improveprediction accuracyVSAvoiddata preparation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent uses image copying and transformation techniques to generate training data. Instead of manually labeling objects, the system creates synthetic training examples by copying images and applying transformations to simulate different robot actions and outcomes, thereby eliminating the need for complex manual data preparation while maintaining prediction accuracy

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system performs self-service by automatically generating its own training data through image transformations and robotic execution of predicted actions. The robot autonomously creates training examples by executing predicted movements and capturing the resulting images, eliminating dependency on manually labeled datasets

Inventive Principle:
Principle #25Self-service

2Ease of operation

If existing video prediction techniques are used for complex physical interactions, then implementation simplicity is maintained, but prediction accuracy deteriorates in real-world scenarios

Engineering Contradiction:
Improveimplementation simplicityVSAvoidprediction accuracy
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The patent replaces traditional mechanical physics-based prediction models with a neural network-based visual prediction system. The neural network learns physical interaction patterns directly from image data, substituting complex mechanical modeling with data-driven learning that achieves higher accuracy while remaining implementationally simple

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The system changes the parameters used for prediction from physics-based mechanical parameters to visual parameters captured in images. By training the neural network to recognize visual patterns of physical interactions, the system achieves accurate prediction of complex interactions without requiring explicit physics models

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If short-term predictions are made for robot actions, then prediction accuracy is improved, but the ability to plan ahead is limited

Engineering Contradiction:
Improveprediction accuracyVSAvoidplanning horizon
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary predictions of multiple future states before the robot actually executes actions. By predicting short-term outcomes repeatedly in sequence, the robot can plan ahead multiple steps, effectively extending its planning horizon while maintaining the accuracy benefits of short-term predictions through iterative refinement

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3689557B1Machine learning methods and apparatus related to predicting motion(s) of object(s) in a robot's environment based on image(s) capturing the object(s) and based on parameter(s) for future robot movement in the environment
Publication Date: 2025.12.03 GOOGLE LLC
  • EP3689557B1 patent drawingFigure 1
  • EP3689557B1 patent drawingFigure 2
  • EP3689557B1 patent drawingFigure 3

AI summary

Some implementations of this specification are directed generally to deep machine learning methods and apparatus related to predicting motion(s) (if any) that will occur to object(s) in an environment of a robot in response to particular movement of the robot in the environment. Some implementations are directed to training a deep neural network model to predict at least one transformation (if any), of an image of a robot's environment, that will occur as a result of implementing at least a portion of a particular movement of the robot in the environment. The trained deep neural network model may predict the transformation based on input that includes the image and a group of robot movement parameters that define the portion of the particular movement.