HUMEMBR: Learning Human Routines for Predictive Embodied Navigation
Samira Huber, Klaas Pelzer, Duc M. Nguyen, Xuesu Xiao, Sören Pirk · 2026
Read HUMEMBR: Learning Human Routines for Predictive Embodied Navigation by Samira Huber, Klaas Pelzer, Duc M. Nguyen, Xuesu Xiao, Sören Pirk (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-30404.txt.
Description
Samira Huber and four colleagues give robots a memory of human routines, not just maps of rooms. People follow daily patterns in homes and offices. They arrive, leave, and linger at predictable times and places. A useful robot should answer where someone likely is. It should also predict when they usually leave. HUMEMBR is built for exactly those questions.
The system constructs memory continuously as the robot observes. A parallel mechanism retrieves and answers queries from that memory. Structured representations of routines accumulate over long periods. Experiments compare it against full-context language model baselines. HUMEMBR reasons better about long-horizon behavior, the results show. It also uses substantially fewer tokens doing so. The team deployed it on a physical robot. Two different real environments hosted the live tests. Diverse queries and navigation tasks were handled successfully.
Catalog record
| CREATOR | Samira Huber, Klaas Pelzer, Duc M. Nguyen, Xuesu Xiao, Sören Pirk |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36 KB |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.30404