Imagine each site surveyed and recorded many times a day, in real-time, with millimetre accuracy, generating up-to-date, contextualised 3D drawings and fully navigable visualisations. Imagine being able to remotely inspect and validate every project in real time, from your tablet or computer, without constantly moving from one site to another. Imagine the ability to virtually rewind every project to a particular point in time, peeling off all the layers of later work so you can see exactly where the plumbing went, or electrical cabling, or if the steel beam has the right padstones below it. Now imagine the software being able to check all of this in real time (compliance with drawings, specification, regulations etc) and flag and alert you of any discrepancy immediately, saving all the rework.
Imagine the impact of this on your and others’ productivity, project costs, project quality, insurance costs, accountability, sustainability and ease of maintenance later on. This is what AIRA’s real-time high-fidelity data acquisition will enable.
Building World Models
Robots can’t be deployed productively without ‘abilities’ which in turn cannot be developed without intricate models of the target environment. Those models will require data from 1,000s of target sites, collected and processed in a way that is useful, intricately defining each action needed to change existing state into the desired state. Millions of simulations later robots will be deployed to sites with heavily pre-trained simulations updated in real time to account for unpredictability. A ‘human in the loop’ will supervise remotely to ensure they don’t get spooked or drift into a hallucination, correcting errors or taking over at times. None of the above has been achieved so far and most of it is impossible with the current state of technology (apart from data collection and processing). The AIRA team came closer than anyone else, deploying semi-autonomous teleoperated robots to construction sites and delivering over 5,000 jobs.
AIRA captures real world environments in dynamic, productive settings, defines each action and states between actions to build accurate World Models and simulate and train the deployment of Physical AI. AIRA will imbed Data Scouts on as many productive sites as possible, across a number of sectors, to continuously collect the design, progress and works data enabling the clear breakdown of all productive actions into defined changes between individual ‘states’, allowing creation and commercialisation of time specific, contextualised Datasets and Physical AI training models.
Hi-fidelity data is vital for training AI systems. Most available data is unstructured, and AI systems find it difficult or even impossible to process. Data needs to be labelled and contextualised, which is complex and resource-intensive, especially with large datasets. Synthetic data can help with some scenarios involving simple, non-deformable external environments but it doesn’t have enough granularity to be useful in detailed, unstructured environments, especially when dynamic changes and interactions with other objects occur.
The AIRA team has been collecting and processing model-usable scans from construction sites, using various technologies, and has built up vast expertise in this field. Ultimately, the purpose of data acquisition, starting with point clouds, repeatedly captured on various sites, with the help of labelling, tokenisation, fundamental computer visualization and graphics transformations, is to provide a gymnasium in which to train Physical AI to do a particular task. Training Physical AI hinges on the availability of vast amounts of high fidelity, context rich data.