01 / DAY 1 · EXPLORE
Find the other elevator.
Explore approximately 87 m from room 625 to the opposite elevator and retain meaningful places along the route.
A little curiosity. A world of experience.
EVOLIFE✦
Ask to Evolve.
A unified harness for lifelong embodied navigation
with hybrid action space and persistent memory.
† Corresponding authors
Tongji University · Shanghai, China
Navigate with intention. Interact with curiosity. Evolve through experience.
EVOLIFE / ONLINEWORLD 01
A new place. A new possibility.
↗
navigate_candidate✦A place becomes a memory.
An interactive illustration. Say hello to the robot.
8Benchmarks
4Tasks
3Decision backbones
InfiniteCuriosity to carry forward
01 / THE IDEA
Give navigation
a little initiative.
A robot should do more than follow the next command. It should choose a useful way forward, ask when something is unclear, and carry that understanding into its next visit.
navigate_candidate / navigate_pixel / navigate_object / stepAgent chooses · Adapter executesA movement intention can be expressed as a geometric candidate, a selected floor pixel, an object estimate, or a short discrete action. Actual observations and motion feedback inform the next decision.
Read the abstract
EVOLIFE connects hybrid actions, active interaction, and persistent spatial memory within one embodied navigation harness. Each encounter can inform the current decision and contribute experience for subsequent tasks.
02 / FIELD NOTES
Small decisions.
Whole worlds.
Explore navigation across eight benchmarks. Watch the views, maps, selected tools, and agent-initiated conversations come together.
03 / OUT INTO THE WORLD
The world changes.
Experience stays.
A corridor becomes familiar. A door closes. A new day begins. Multi-day physical navigation brings action, help-seeking, and remembered places into the same ongoing experience.
02 / DAY 1 · ADAPT
A door changes the plan.
Ask for help, verify the open passage, and continue to the remembered destination.
03 / DAY 2 · REVISIT
A familiar place, revisited.
Restore the spatial context and navigate approximately 60 m to the remembered glass office on the following morning.
04 / THE EVIDENCE
Across tasks.
Across encounters.
Eight benchmarks span language-guided routes, object search, interactive instance navigation, and continuous multi-task sequences.
SR on first visits to unseen environments: R2R-CE, RxR-CE, HM3D-OVON and HM3D v2.
82.23↗93.31%
SPL on matched revisits without new interaction.
LHPR-VLN / GOAT-Bench, 50 full sequences each.
THE COMPARISON TABLESOne harness.
One harness.
Eight benchmarks.
Loading comparison results…
A CLOSER LOOK AT THE NUMBERS
Experience, efficiency, and the decision backbone.
05 / A CLOSER LOOK
Behind the next move.
EIGHT BENCHMARKS / EIGHT ENCOUNTERS
Decisions, up close.
Read the task. Follow the growing map.
See how each observation becomes a tool choice.
Loading cases…
Loading the eight qualitative examples…
TAKE THE NEXT STEP
Stay curious.
Keep exploring.
A new encounter. A better way forward.
Read the paper on arXiv. Code is coming soon.
BIBTEX
@misc{sheng2026askevolveunifiedharness,
title = {Ask to Evolve: A Unified Harness for Lifelong
Embodied Navigation with Hybrid Action Space
and Persistent Memory},
author = {Kai Sheng and Liuyi Wang and Jinlong Li and
Haojie Dai and Chengju Liu and Qijun Chen},
year = {2026},
eprint = {2610.39446},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.39446}
}
