AI learning of arbitrary plants
We use deep neural networks to teach the robot arbitrary new plants. You just put a few examples on the conveyer belt, use our RVAI portal to create a model, and within few minutes you have a robot planting a new plant type. Next time you buy flowers for your loved one, think about our deep learning engine.
On the video you see an installation in Carleton, Michigan. The video at the bottom has a four gripper version of the system.
Deep Learning in a car manufacturing plant
To assemble a car, a car manufacturer needs a solid and validated flow of components. High tech scanning technology can guarantee that a specific component will arrive just in time on the production line. Audi relies on Robovision to fulfill this need.
By applying 3D deep learning on manufactured pieces Robovision checks if the right subtype is selected, and if a set of components is complete for a specific car type. In this case RVAI ensures that the WZB (the kit you use when you have a flat tyre) is complete and validated. Thanks to the strong GPUs from NVIDIA we keep the cycle time fast and snappy.
Deep learning in the packaging industry
Pepsico manufactures and packages dorito chips in the highlands above Mexico City. Package branding is changing so frequently changing that one needs a powerful AI based central system to teach in new branding very fast. RVAI is the solution here. Reflective, blown up and unpredictable, recognizing branding is not for the faint hearted in this application.
AI applied to 3D in Agriculture
The next horizon in our agricultural developments is the art of convolutional networks applied to 3D. Now every day a Jumbo 747 is arriving from Africa full of Chrysanthemum stems. We want to grow them here in Western Europe and manipulate them with smart robots (that pick the stems from the mother plant. In this way we can use the fertile ground of Ethiopia to feed the hungry in the region. The AI revolution can generate powerful secondary effects, disrupting the flow to cheap labour and create a more harmonious world order.
Deep learning applied to the process industry
We apply deep learning to processes. By mounting a series of in house developed embedded sensors on machinery and piping (http://www.altifort-pvi.com/ ), we are able to generate data from small variations in the system that is used as input for our RVAI. In this way we can teach RVAI what normal stationary behaviour is. In case of anomalies, it is detected properly. The advantage is: the more data, the better it works. In this way we are creating the valves of the future, that know when and how they are malfunctioning.
Deep learning applied to the electronics industry
We apply deep learning to detect quality issues in electronic devices. Our iPhones and Samsung phones need to be ultrarobust. Nobody wants an unstable phone and advanced AI techniques can make the manufacturing process more robust.
AI applied to the surface inspection done by scanning electron microscopy
We applied machine learning algorithms to inspect surfaces that underwent a special nanotechnological surface treatment. The degree of corrosivity is essential in the subsequent process, and the quality of the final product is highly depending on the exact degree of corrosivity.
AI applied to thermal streams
One of the biggest chemical sites in Europe (Chemelot) is relying on Robovision to detect ammonia leaks while pumping over ammonia to the site tanks. Our Gashawk™ software is using state of the art algorithms to distinguish between endothermal reactions related to ammonia, and other similar phenomena.
For more compact projects, we use embedded technology to run AI on Flir cameras. With techniques such as squeezenet and noscope, we optimize complex architectures for low power devices. ThermalHawk™ is the platform we offer here. Below also a picture where we performed process optimization at a Polish plant with thermal data.
Multimodal deep learning applied to xray
One of the main advantages of deep learning is the amount of different complex data that can be aggregated and used as input for the deep architecture. Combining hyperspectral imaging, 3D point clouds and xray can generate a lot of operational interesting data, and interesting snippets of information can be combined in an intermodal way (see more in this video).
In the video below we present you a case of deep learning based morphological segmentation of seeds (Gerbil). At the bottom of the page, we present the Ixcon use case, done in conjunction with IMEC, where we want to achieve rapid inline xray segmentation and quality control, without performing the full CT scan as depicted.
Deep learning based avatar search in urban environments
(due to NDA and pricacy reasons this video is intentionally blurred)
Our deep learning engine is connected to the Genetec platform, we can analyse security streams in realtime and look for people with specific traits and clothing. Imagine an missing child in an urban environment that needs instant action.
Thanks to a powerful IBM backbone this search can be performed in realtime.
Our software is used to run the fully automated pharmacies of the future. We drive the image processing behind the scenes, and recognize the medical boxes. Medical bottles, boxes, everything can be processed.
Automated creation of UDP
Together with R.I.S. we developed machinery to create unit-dose packaging solutions. A rig of cameras detects pills of all kinds and the robot glues them on a strip. The result is used in high tech hospitals.
Vision and image processing enables fully automated warehousing. Here our systems are critical for safety, recognizing the type of stones, and the size of the palette before it goes into the ‘Manhattan’ warehouse. Xella is relying on Robovision for this kind of critical, image processing based use cases.
Image processing in the food industry
Currently we are combining hyperspectral imaging and deep learning in the food processing industry, these tracks are still in progress (see the explanatory images below about the principles behind hyperspectral imaging)
Regarding the video at the bottom of the page: this is an older project (2011) where more usual image processing is applied (image segmentation). Nevertheless we show it here to demonstrate our ability to efficiently build prototypes for different industry niches.
Many more case stories:
- We are using deep architectures to get to the bottom of hyperspectral imaging.
- We are using a special brand of convolutional neural networks (LSTM networks) to detect anomalies in cyber security related logfiles (for now with limited success, still a long way ahead, images and video have no secrets for us, text files however…).
- We are looking into FPGA and embedded development to keep our edge and to remain in the blue ocean part of this rapidly emerging new business.
- We applied eulerian video magnification to read heartbeats out of normal video streams.