A robot assembling a toy truck it has never seen before does better when you just tell it the rules in plain English.
Researchers built a neurosymbolic system, part neural network, part symbolic reasoner, that learns to put together novel structures from unfamiliar parts under constraints it never saw during training. The setup: an agent working in a simulated toy-truck-assembly domain that, after deployment, runs into semantic rules it was not trained on, like "dump trucks have a dumper." It has to pick up both the missing vocabulary and the missing rules from user interactions while it works, drawing on natural-language hints and dense visual observations of the parts in front of it. In testing, the team compared ways of teaching the agent: naming parts alone, task demonstrations alone, and explicit constraint language. Constraint language won, cutting the amount of data needed for the agent to adapt on the fly.
That is the useful bit for anyone building assembly robots, warehouse pickers, or any embodied agent meant to work with parts catalogs that change after training ends. Retraining a model every time a supplier swaps a part is expensive; an agent that can absorb "this part goes here because trucks need one" from a sentence is cheaper and faster to update. It is also a small data point against the assumption that more demonstrations always beat more explanation.
Worth remembering this is a simulated toy truck, not a factory floor. Real hardware, real sensors, and real human instructions are messier than anything in a lab domain, and the gap between learning fast in simulation and working on a warehouse line has swallowed plenty of promising robotics papers before.