"No worries" at the wrong time: three years of social robot under LLM deliver five design lessons

The robot Alice said the right words in the wrong order, or announced a visitor already present. A feedback offers TRACE, a review to identify these flaws.

When a person declined their invitation to play, the robot Alice would turn away, say "No worries, it's all good", and then leave only after a pause: each element was plausible, but their sequence changed the message. Another time, he warned a host of the arrival of his visitor... who was already standing next to him. These episodes come from a three-year consortium around ROS 2, which integrated the major language and vision models as they appeared.

Eliahu Khalastchi draws a central idea from this: foundation models expand what robots perceive and can do, but this does not guarantee that a change in situation will have an impact on the action ultimately addressed to the person. Information may never be collected, a constraint may be lost between planning and execution, and the system may even declare success when the action no longer serves its purpose.

The article formulates five lessons (attention, knowledge, commitments, body synchronization, authority over action) and proposes TRACE, a light review that modifies an interaction hypothesis and follows it to the decision point. Applied to a published system of behavior trees generated by LLM, it shows that an available clarification mechanism can remain unused. The text is submitted to IEEE Robotics & Automation Magazine.