Retrocausal named to the 2022 CB Insights' Advanced Manufacturing 50 List
How It Works
Learn about two ways of rapidly setting up Pathfinder on a new process and how flexibly a process can be modified to incorporate changes.
Recording a few shifts for holistic tracking
Pathfinder requires the user, e.g. a quality engineer or a production lead to define the standard operating procedure for the process. Next, the user sets up the software to record five or more shifts of the assembly or kit packing process through the Pathfinder web interface. The advantage of capturing multiple shifts to train our machine learning models is that they adapt to slight variations in performing the process. As soon as these shifts have been recorded, Pathfinder sends a notification to the user, and requests them to "label" one cycle of the process.
The labeling is also performed using the Pathfinder web interface, and is in the form of time series labels, i.e. the user provides associations between step names and start-stop times for the steps. The labeling process usually takes about ten minutes of the user's time. Once the labeling process is complete, the user presses the "Request Training" button on the web portal, and the data is sent for smart annotation and training. Usually it takes 36-48 hours to get a trained model back, if the process is performed in a standardized manner.
In contrast, to IoT-based solutions or hand-tracking only solution which only digitize 20% of steps in an average process, this Pathfinder workflow is able to capture 90% or more of the steps in a process with negligible false-positive rate.
Hand-targets for approximate tracking from a single frame!
In some cases, engineers and managers are content with relatively coarse-grained tracking of only a subset of steps. For these cases, we allow users to simply associate spatial target regions in a single camera image with steps in the Standard Operating Procedure. At run time, if the operator's hand touches the target region corresponding to a step, that step is considered as "performed". This alternative mechanism for setting up tasks requires five minutes of work from the user in total, and doesn't require "training" any new machine learning model, which simplifies deployment.
If the unit being assembled is considerably larger than a human hand, and steps take place at distinct locations on the unit with parts picked from unambiguous locations, the approach works well. However, if the unit being assembled is compact and there's considerable tool usage, the approach may not capture some steps robustly because hand location alone is a weak predictor of step performance.
Flexible Step Manipulation and Verification
Pathfinder allows conveniently modifying processes for which data has already been collected. It is possible to add, delete, rename, change ordering as well as merge step definitions with a few clicks. Pathfinder also enables steps to be marked as "optional" which means that operators aren't alerted if those steps are missed.
Pathfinder also provides the option to explicitly define the end-states of steps by specifying visual appearance of the unit upon successful step completion. This allows verifying that a step (or group of steps) achieved their desired goal.
Schedule a demo
Get access to the leading video analytics platform for your factory floor.