Opinion
In Canoa, the canvas is dimensionally unlimited (all images sourced from US Patent Office)
The press release headline promised, “The Future is Here: AI-Powered Interiors by Canoa.” The text of the release explained what patent #11,977,821 granted Canoa:
“The patent is the first to embed a learning algorithm directly inside of a CAD application and covers the generation of user-specific recommendations using their own clicks as prompt inputs.”
I was intrigued, and set about to read the 52 pages of the patent, so that you don’t have to. It’s available through https://patents.google.com/patent/US11977821B1/.
The thing about patents is that their text is repetitive to the extreme, so that there can be no vagueness as to what is meant. So, for instance, the phrase “The CAD design and object acquisition and management computing system 100” appears over and over -- 74 times. It’s talking about the Canoa software, but it can’t just say “It.” Another repetitive phrase, “In some aspects,” appears 129 times. It’s a daunting read.
One other thing: a patent covers a “method,” a series of steps of accomplishing something, and so that word appears a lot, as well. (This is why ideas cannot be patented.)
What The Patent Appears To Cover
This patent appears to cover a method that “automatically” suggests furnishings that would fit an area of a specific floor plan, based a several criteria. Furnishings (symbols, parts, blocks, components) are sourced locally or remotely from servers hosted by a variety of product manufacturers and second-hand sellers, along with data useful to designers like availability and delivery dates.
There is a suggestion that Lidar scan data could be input from remote locations to size the rooms. Other input is from existing drawings. The output is 2D and 3D drawings, and a parts list of selected furnishings (aka shopping cart).
The patent spends much time discussing methods already common to CAD and other graphical systems, such as dragging’n dropping furnishings in the floor plan. What might be new is having existing objects move to accommodate newly-added ones.
Other methods not new to us include the CAD system being locally installed or running on the cloud; adding markups and comments; accessing block libraries; sharing floorplans simultaneously or by links; recording object lifecycle data (important for tracking items through second-hand shops); determining environmental impacts, including transportation of products; assigning unique identifiers (UIDs) to objects in drawings; and undertaking facilities management. A curation module locates all furnishings in a group with each other. The entire group can then be drag’n dropped into the floorplan. I don’t know why the patent had to include all of these “prior arts,” as they are known in the patent world.
Finally, AI
Artificial intelligence got top billing in the press release, but it isn’t mentioned until page 19 (line 45), where the patent suggests that ML/AI [machine learning, artificial intelligence] could be used to suggest kits (items) to complete the design. How this works is not detailed.
Later, the patent refers to rules-based and classification machine-learning, which possibly might be Canoa’s definition of AI. Still later, the patent suggests that generative AI could be used to recognize patterns.
Page 31 finally describes in some detail how rules-based modeling is used to place furnishings, based on parameters like lighting, minimum distances, and relationships (tables with chairs). The rules proposed by the patent include the following:
- Impact of a recommended object
- Cost
- Minimum number of shared properties between an object and a recommended object
- Relationships between a first and second object (height, etc.)
- Design preferences for particular organizations (color palate, texture limitations/preferences, style preferences, and the like)
- Minimum object rating requirements
- Other rules suitable to selecting and generating a recommendation
The recommended objects can be further refined by
- Past group purchases from other organizations
- Design ratings provided by system users
- An internal set of design rules
- Physical item properties (chair height, etc.)
"Training" the AI is described on page 37 (line 18), where users can give feedback on the accuracy of the prediction, approving or correcting the object’s type, color, condition, model, and so on.
In the end, a specific AI system is never specified, only suggested, such as on page 35 (line 5): “...the machine learning model may be configured using a variety of different types of supervised or unsupervised trained models such as, for example, support vector machine, naïve Bayes, decision tree, neural network, and/or the like.”
Yes, We Have No IFCs
There is no mention of using IFC data to classify objects.
Instead, page 22 (line 30) starts describing a multi-level classification system for identifying furnishings and their makeup based on BERT (bi-directional encoder representations from transformers), which is normally used in natural language processing.
I may have missed it, but there is no explanation of how the method reconciles different classification systems of incoming furnishings. I would think IFC would be the most common one, so proposing a different system from the one used by the CAD world seems to me somewhat odd.
What Ralph Grabowski Thinks
In a rush to be AI-hip, some other CAD vendors are redefining older technologies as “AI,” such as generative design. The patent is more subdued about the use of AI than is the press release.
This patent seems to me to define “AI” as rules-based, which I do not consider a form of artificial intelligence. Rules-based is a series of pre-programmed if-then statements. In my opinion, no thinking by an artificial entity is involved.
Similarly, allowing the user to adjust the objects selected by the system is also not AI, in my opinion: it is filtering. Just as we ask a spell checker to add a new word to its list of recognized words, the filters create lists of acceptable objects.
In the end, I could find no description in the patent of what I would consider an AI system. If I got it wrong, I am happy to be corrected.
Comments