A robot's footage helped locate a forgotten jacket. AgiBot opens its continuous-observation tool.

Where did that jacket go? AgiBot's LIFE system traced it to the back of a pantry chair in a day's recording and retrieved the corresponding image. The idea is to turn earlier observations into evidence for questions that arrive later.

On November 10, the team added real-time RoboBeats code and a local web demo to the research first posted October 30. Developers can connect a camera, load video and specify conditions to watch for. The tool samples once per second and reuses computation within a video chunk. Rather than generating a description for every frame, it scores whether an event has happened. The code uses Apache 2.0 licensing.

Think of this component as a filter before memory: it selects moments around a detected change for later processing. Three model sizes give developers hardware options. The immediate offering is a way to experiment with continuous perception, not a finished household lost-and-found robot. Chinese technology outlet MyDrivers covered the framework on November 11.

The paper's day-long memory study used one robot and six questions. Memory was built offline from recorded footage, so it does not establish long-term autonomous operation. The real-time tool has its own caveat: the authors have not benchmarked accuracy on users' custom conditions, and wording matters.

Source times: LIFE paper revision, November 4, 2026, 21:26:08 Beijing / November 4, 08:26:08 EST (first submission: October 30, 23:23:52 Beijing / October 30, 11:23:52 EDT); RoboBeats real-time code, November 11, 00:05:41 Beijing / November 10, 11:05:41 EST; MyDrivers, November 11, 16:37:52 Beijing / November 11, 03:37:52 EST.