Periodic Reporting for period 1 - SeConRob (Self-configuring multi-step robotic work-flows)
Reporting period: 2023-10-01 to 2025-09-30
The SECONROB project aims at developing methods for the self-configuration of robotic processes, where each manufacturing step depends on the results of the previous step. In this case a lot of productivity, energy and resources are lost, because the processes currently cannot be automated for technical and economic reasons.
Objective - O1
To deduce models of a range of processes (inspection, welding, grinding and polishing) that can be automatically configured and executed on a robot in the context of a multi-stage manufacturing process. This will be based on previous projects, where abstract, quite generic models have already been developed, e.g. for inspection processes.
Objective – O2
To develop data analysis methods that convert raw data from the upstream process into features that are relevant for the configuration of the downstream processes. AI will be applied for the raw data analysis.
Objective – O3
To implement methods for the automatic generation of robot programs for the downstream processes, including a long-term feedback loop based on reinforcement learning that improves the quality of the resulting processes.
Objective - O1
To deduce models of a range of processes (inspection, welding, grinding and polishing) that can be automatically configured and executed on a robot in the context of a multi-stage manufacturing process. This will be based on previous projects, where abstract, quite generic models have already been developed, e.g. for inspection processes.
Objective – O2
To develop data analysis methods that convert raw data from the upstream process into features that are relevant for the configuration of the downstream processes. AI will be applied for the raw data analysis.
Objective – O3
To implement methods for the automatic generation of robot programs for the downstream processes, including a long-term feedback loop based on reinforcement learning that improves the quality of the resulting processes.
In the project two usecase are considered to achieve the expected outcomes. The usecases deal with the setup of multistage robotic processes with adherent process steps. Use case 1 addresses the manufacturing of safety-critical metallic parts produced by Safe Metal SA. Use case 2 addresses the inspection of forged aerospace components manufactured by Otto Fuchs.
Expected outcomes:
- Parametric process models will be created that can be configured automatically through the proper selection of their parameters and that do not require programming effort for the realization of an individual process (inspection and re-work processes).
- Robust data analysis based on AI methods will be made available for manufacturing environments. Synthesis and augmentation of training data, which are inevitable in this domain, will use realistic models to enhance the data sets, minimizing the “reality” gap and making sure that AI methods can deal with the usual variation expected in this industrial domain.
-The feedforward transfer of knowledge generated at the upstream process step will be used seamlessly and fully automatically for the planning for the downstream process. No human intervention shall be needed. To optimize the integrated, multi-stage process over a longer period of time, reinforcement learning methods will be developed that provide feedback from the results of the downstream process to the upstream process. The feedback can be either from human experts or automatic.
Expected outcomes:
- Parametric process models will be created that can be configured automatically through the proper selection of their parameters and that do not require programming effort for the realization of an individual process (inspection and re-work processes).
- Robust data analysis based on AI methods will be made available for manufacturing environments. Synthesis and augmentation of training data, which are inevitable in this domain, will use realistic models to enhance the data sets, minimizing the “reality” gap and making sure that AI methods can deal with the usual variation expected in this industrial domain.
-The feedforward transfer of knowledge generated at the upstream process step will be used seamlessly and fully automatically for the planning for the downstream process. No human intervention shall be needed. To optimize the integrated, multi-stage process over a longer period of time, reinforcement learning methods will be developed that provide feedback from the results of the downstream process to the upstream process. The feedback can be either from human experts or automatic.
O1: Modelling of processes to enable automatic configuration
SECONROB combines the challenges of two path planning methods (covering a certain area and autonomous real-timeplanning): the path planning processes are comparably complex, requiring full coverage with high certainty and quality, while having the need for autonomous, (near) real-time planning without human intervention. Through the reactive coverage planning methods, that use a process model and have the ability to quickly adjust a pre-planned path to the actual geometry of the part or to the area that needs to be processed, SECONROB will achieve progress beyond the state of the art and will enable the self-configuration of robotic processes.
O2: AI-based data analysis for NDT processes to extract re-work information
In SECONROB three NDT modalities will be investigated and analysed by AI-based methods in WP4. The data involved comes from magnetic particle inspection, eddy current testing and ultrasound testing. The individual steps of the AI-system are to detect, segment and classify defects of each inspection method and characterize them in a way that enables following robotic downstream processes to self-configure and optimize themselves. For all these steps the data analysis within SECONROB will focus on Deep Learning methods which have been proven to outperform traditional methods in many machine vision tasks.
O3: Methods for automatic optimization of robot programs
In SECONROB the task to be executed by the robot will consist of a trajectory and process parameters to be applied along this trajectory. Previous work was focused on generating a robotic trajectory and process from scratch using reinforcement learning. Due to the limited training input, the training process obviously takes long and is difficult to control. It might lead to feasible, but ‘strange’ solutions. Within WP5 of SECONROB a new approach is planned, that differs in two main aspects:
• There is a lot of knowledge available about the processes that the robot should do. This includes the physical understanding of the process as well as expert knowledge that has been acquired in industrial practice. Within SECONROB such expert and physics-based knowledge will be used to calculate an initial trajectory/process for the robot. It is expected that this initial process will already be a good solution for many cases. The reinforcement learning will be used to further tune and optimize it and to capture all the properties that are not modelled in the physical process model.
• SECONROB combines two reinforcement learning approaches, the offline RL to enable learning control policies by utilizing only prior experience, before the process is put into operation. To evaluate the learned policy, we concatenate offline RL with the interactive RL approach, where the training input for the interactive reinforcement learning will be provided by an expert or as some high-level information coming from an automatic system (such as accept/reject from the end-of-line test).
It should be noted that the interactive reinforcement learning process takes place over a large number of parts, i.e. it does not aim at improving the process results for a single part, but to enhance the implicit understanding of the process ‘in general’. It will thus lead to an average quality improvement and a reduction of reject parts over time.
SECONROB combines the challenges of two path planning methods (covering a certain area and autonomous real-timeplanning): the path planning processes are comparably complex, requiring full coverage with high certainty and quality, while having the need for autonomous, (near) real-time planning without human intervention. Through the reactive coverage planning methods, that use a process model and have the ability to quickly adjust a pre-planned path to the actual geometry of the part or to the area that needs to be processed, SECONROB will achieve progress beyond the state of the art and will enable the self-configuration of robotic processes.
O2: AI-based data analysis for NDT processes to extract re-work information
In SECONROB three NDT modalities will be investigated and analysed by AI-based methods in WP4. The data involved comes from magnetic particle inspection, eddy current testing and ultrasound testing. The individual steps of the AI-system are to detect, segment and classify defects of each inspection method and characterize them in a way that enables following robotic downstream processes to self-configure and optimize themselves. For all these steps the data analysis within SECONROB will focus on Deep Learning methods which have been proven to outperform traditional methods in many machine vision tasks.
O3: Methods for automatic optimization of robot programs
In SECONROB the task to be executed by the robot will consist of a trajectory and process parameters to be applied along this trajectory. Previous work was focused on generating a robotic trajectory and process from scratch using reinforcement learning. Due to the limited training input, the training process obviously takes long and is difficult to control. It might lead to feasible, but ‘strange’ solutions. Within WP5 of SECONROB a new approach is planned, that differs in two main aspects:
• There is a lot of knowledge available about the processes that the robot should do. This includes the physical understanding of the process as well as expert knowledge that has been acquired in industrial practice. Within SECONROB such expert and physics-based knowledge will be used to calculate an initial trajectory/process for the robot. It is expected that this initial process will already be a good solution for many cases. The reinforcement learning will be used to further tune and optimize it and to capture all the properties that are not modelled in the physical process model.
• SECONROB combines two reinforcement learning approaches, the offline RL to enable learning control policies by utilizing only prior experience, before the process is put into operation. To evaluate the learned policy, we concatenate offline RL with the interactive RL approach, where the training input for the interactive reinforcement learning will be provided by an expert or as some high-level information coming from an automatic system (such as accept/reject from the end-of-line test).
It should be noted that the interactive reinforcement learning process takes place over a large number of parts, i.e. it does not aim at improving the process results for a single part, but to enhance the implicit understanding of the process ‘in general’. It will thus lead to an average quality improvement and a reduction of reject parts over time.