Robots Need to Remember Too: Embodied Memory Competition Through the Lens of an IROS Best Paper

If a robot has just put the remote control in a drawer and a few minutes later you ask it to fetch the remote, the first problem it faces may not be "how to grasp it" but "where is the remote." The answer is not in the current view; it exists only in the past.

This is precisely the starting point of robot memory research: current perception does not always fully describe the current world. A robot capable of long-term operation must let the past participate in present decisions.

The recently concluded IROS 2026 awarded its Best Paper Award to LT-Mem from DGIST APRL. What it studies is not cramming more historical frames into a single task, but how a robot, when repeatedly returning to the same yet constantly changing environment, remembers where objects are now, where they were in the past, and how the world usually changes.

Figure 1 | LT-Mem receives the IROS 2026 Best Paper Award

Putting the work of the past two years together, the changes roughly unfold along four questions: Which tasks truly require memory? What information is worth writing into memory? In what form should memory exist? After the world changes, how should old memories be revised? This article attempts to connect benchmarks, policies, and long-term scene understanding along a progressive line.