ASENA · 2026

Self-evolving agents for embodied navigation

ASENA turns execution feedback into reusable navigation skills. A coding agent composes, tests, and revises its programs, with tools for navigation in simulation and whole-body action on a humanoid.

Navigation is an open-ended programming problem: the next instruction, unfamiliar room, or failed route can demand a behavior the agent has not written yet. ASENA uses a coding agent to build and refine navigation programs from execution feedback. Observations, tool results, and task outcomes inform changes to code and reusable skills, so experience can carry into subsequent attempts. We study this self-evolution process in R2R and RxR simulation, and extend the framework to a Unitree G1 for navigation, upper-body gestures, and speech. A trained VLN policy provides a useful navigation tool in simulation; its architecture and curated training data support the agent alongside perception, memory, and planning tools.

  • Self-evolving navigation: a coding agent inspects execution feedback, revises programs, and accumulates reusable skills.
  • Measured R2R and RxR training curves on recurring tasks, with recorded replay-skill usage and a documented protocol.
  • A whole-body extension on the G1, plus a trained VLN tool and curated atomic-navigation data for simulation.

Demos

Real-world Demo

The G1 demonstrations show navigation as part of a larger task: explore a space, gather evidence, return to a person, and report what was found. Whole-body control extends the agent’s action space to movement and gestures, with speech as another output.

No pre-built maps. Each map is reconstructed online as the robot explores.

Embodied reasoning · Self-evolution in a home

Explore the world.
Write what comes next.

The answer is somewhere in the house.
Every quest can create, refine, or reuse a skill.

THE MISSION

Is the mirror closer to the floor than the window with curtains?

SKILLS REUSEDfloor_change · height_above_floor

Navigate down from the attic with a repaired floor skill, then reuse a prior height skill to compare both targets against the floor.

Recorded task replay. Both skills predate this quest: floor-change repair from Quests 11/278, height measurement from Quest 67. Selected scan surfaces are restored.

Method

Experience becomes the next skill

ASENACoding agent · Claude Code or Codex
from asena.device import Robot
    Reusable skill library
    into the next pass ↗

    Method · Real world

    From navigation to whole-body action

      SONIC controls 29 body joints: 12 in the legs, 3 in the waist, and 14 in the arms. Agent-authored upper-body trajectories pass through motion validation before execution. The 14 DEX3 finger joints and onboard speech are separate from SONIC’s body control.

      Method · Simulation

      Navigation skills in simulation

      A learned navigation tool

        Curated atomic-navigation data Browse navigation examples

        Quantitative results · Simulation

        Learning across repeated tasks

        Full measurements, protocol & data

        Standalone navigation tool

        Citation

        BibTeX

        Project-page citation. The paper citation will be added with the manuscript release.

        @misc{asena2026project,
  title = {ASENA: Self-Evolving Agents for Embodied Navigation},
  url   = {https://asena-bot.github.io/},
  year  = {2026},
  note  = {Project page; paper citation forthcoming}
}

        Acknowledgements

        We thank Jarred Travers, Amanpreet Singh, Peter Pham, and Xiangchen Tian for their help with real robot infrastructure.