Robot Grasp Policy Training With Pseudo-Labels for Unseen Objects

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing robotic manipulation techniques face challenges in adapting to unseen objects and new environmental conditions due to their reliance on supervised learning and offline training, which results in inefficiencies in online grasp learning.

Innovation Solution

A method for training a control policy using a semi-supervised learning scheme, which generates augmentations and pseudo-labels for image data, allowing for efficient online grasp learning by leveraging unlabeled data and addressing sparse reward feedback.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If supervised learning and offline training are used, then the model can be trained with labeled data, but the ability to adapt to unseen objects and new environmental conditions is limited

Engineering Contradiction:
Improveadaptability to unseen objectsVSAvoidonline adaptation capability
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system performs preliminary offline training with labeled data to establish a baseline control policy, then prepares for online adaptation by implementing a semi-supervised learning framework that can continuously learn from new unlabeled data encountered during operation

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system transitions from static offline training to dynamic online learning, where the control policy continuously adapts to new objects and environmental conditions through semi-supervised learning that combines labeled and unlabeled data streams in real-time

Inventive Principle:
Principle #15Dynamics

2Quantity of substance

If only labeled data is used for training, then the training process is straightforward, but the amount of available training data is insufficient leading to sparse reward feedback

Engineering Contradiction:
Improveamount of training dataVSAvoidsparse reward feedback
Core Design Contradiction:
Quantity of substanceVSLoss of information

Solution Approach 1:

The system introduces pseudo-labels as an intermediary mechanism that bridges labeled and unlabeled data, allowing unlabeled data to contribute to training through generated pseudo-labels while maintaining the benefits of supervised learning signals from labeled data

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system changes the labeling parameter from binary (labeled/unlabeled) to a continuous spectrum by generating pseudo-labels with varying confidence scores, allowing flexible utilization of unlabeled data while maintaining training quality

Inventive Principle:
Principle #35Parameter changes

3Productivity

If semi-supervised learning is used to leverage unlabeled data, then learning efficiency is improved, but the complexity of the training process increases

Engineering Contradiction:
Improvelearning efficiencyVSAvoidtraining process complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The training process is segmented into distinct phases: offline pre-training with labeled data, online pseudo-label generation from unlabeled data, and iterative refinement combining both data types, allowing manageable complexity at each stage while achieving high overall efficiency

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4570450A1Device and method for training a control policy for manipulating an object
Publication Date: 2025.06.18 ROBERT BOSCH GMBH
  • EP4570450A1 patent drawingFigure 1
  • EP4570450A1 patent drawingFigure 2
  • EP4570450A1 patent drawingFigure 3

AI summary

According to various embodiments, a method for training a control policy for manipulating an object is provided , comprising, for each of one or more objects in each of one or more scenes, receiving an input data element including image data representing a shape of the object to be manipulated and its position in the scene, generating, for each input data element, one or more training data elements by generating augmentations of the image data and pseudo-labels for the augmented image data according to a semi-supervised learning scheme and training the control policy using the generated training data elements.