Research
Robot vision, simulated picking, and the business of robotics. We publish the limits alongside the results.
Ultra, a Brooklyn robotics company, just raised a $50 million Series A. Its robots pack real warehouse orders, and customers pay every month. Here is what a working robot business looks like, and what we take from it.
We built an interactive cost model for the Retail Lab. Its default case reaches payback in 33 months. Change one assumption and it takes 65. That sensitivity is the point.
OpenAI put the harness behind Codex into a public beta API. The runtime problem is getting solved for you. The process problem is not.
Adecco is putting an agent into the daily workflow of 27,000 staff. The detail worth copying is not the scale. It is that there is one entry point instead of dozens.
Teradata shipped a layer that plans before it calls a model. The interesting part is what it admits about how agents behave when nobody is watching the meter.
A robot arm now picks groceries out of a messy tote in simulation, choosing what to grab from a camera alone. It works 59% of the time. Getting there meant killing three of our own explanations.
Read: We taught a simulated robot to pick, and were wrong three times first
We rendered 4,600 pictures of a store that does not exist yet, trained a model to find products in them, and got a perfect score. Here is what that score tells us, and what it does not.
Analysts and customers at UiPath Fusion said the same thing in different words. The demo was never the hard part. Getting the undocumented judgment out of people's heads is.
UiPath spent its September conference arguing that the hard part of enterprise AI is the map, not the model. That matches what we see inside real workflows.
It pays off in work that repeats with variation. The person keeps the judgment.
Collaborate with the lab.
We want to hear from robotics engineers, operators, and potential pilot hosts. Bring a real task and the constraints that matter.
Get in touch