ASENA · 2026

ASENA: Self-evolving agents for embodied navigation

ASENA brings coding agents into the physical world. Agents write robot programs, inspect execution feedback, and refine reusable code and skills, so later attempts can build on earlier experience.

Abstract

ASENA is an embodied agent system that connects general-purpose coding agents to robot sensing, computation, supervised execution, and persistent experience. Agents write and execute programs, inspect outcomes, repair failures, and retain reusable notes and skills while keeping their model weights fixed. ASENA-VLN, an optional 4B monocular policy, is trained on route instructions, visual question answering, and curated atomic-navigation data. We evaluate standalone navigation, object-goal navigation, coding-agent navigation, and embodied question answering. Repeated simulation tasks show how retained experience improves success and efficiency. Supervised G1 demonstrations combine search, visual inspection, spatial reasoning, and synthesized gestures without a pre-built map.

An embodied agent system for sensor access, code execution, supervised robot control, and persistent experience.
ASENA-VLN: an optional 4B monocular navigation policy trained on route following, visual question answering, and nine atomic-navigation capabilities.
Evaluation across navigation and embodied question answering, workspace evolution with fixed model weights, and real-world G1 applications.

Demos

Real-world demonstrations

ASENA lets coding agents write and refine programs for perception, navigation, whole-body motion, and simple manipulation through grasp and motion planning. The supervised G1 demonstrations combine exploration and evidence gathering with object pickup and handoff.

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

Embodied question answering

Skill creation and reuse

Simulation examples of skill reuse, creation, and post-task reflection.

Loading the simulation replay…

Watch the recorded episodes: Reuse existing skills, Create a comparison skill, Extract a reusable skill.

Method

Robot harness

A coding agent reads the robot’s observations, composes navigation and manipulation tools, and retains useful code and experience. The robot harness validates motion proposals and obtains operator approval before execution.

Illustrative G1 model; current hardware uses YAM grippers

Evolution in simulation

Sixteen agents run simulated tasks from private copies of a shared workspace. They inspect execution feedback, develop and test skills, and share tips for later episodes. Between rounds, consolidation merges useful discoveries and prunes redundant code. Repeated tasks use the updated workspace; model and policy weights remain fixed.

Canonical workspace v_t↓16 parallel workers

Worker Aprivate copy

Cross the floor patch

skills/onsurface.py

1patch = find_floor_patch(views)

2face(patch.bearing)

3step_and_observe()

4if patch_moved_front_to_back():

5 suggest_stop()

⌘

onsurface.pyRetained in workspace

↗

Worker Bprivate copy

Measure forward progress

skills/doorway_wiggle.py

1start = pose()

2try_local_recovery()

3delta = position() - start.position

4forward = project(delta, start.yaw)

5report_crossing_progress(forward)

⌘

doorway_wiggle.pyRetained in workspace

↗

Worker Cprivate copy

Follow a curved stairway

skills/climb.py

1while step_budget_left():

2 advance_and_read_height()

3 if motion_stalled():

4 probe_alternate_headings()

5 if landing_detected():

6 break

⌘

climb.pyRetained in workspace

↗

⑂

Merge → test → prune

Executable skills · provenance · outcomes

v_t+1↶

Next round: seed every private workspace from the consolidated library

↗

Shared bulletin

STOPPING onsurface.py

A front-to-back floor-color change can indicate crossing a small patch.

RECOVERY climb.py

Probe a new heading when the simulated stair route bends.

↓ Next episode briefing

Read peers’ tips.
Inspect and reuse their skills.

ASENA-VLN

ASENA-VLN is an optional 4B monocular navigation policy that maps instructions and RGB history to waypoints. It is evaluated in simulation and can be called by the coding agent as a navigation tool.

Earlier front-camera observation from a training trajectory

Middle front-camera observation from the same trajectory

Atomic navigation dataset

Our curated dataset targets distinct atomic capabilities, from precise motion and alignment to room goals, door traversal, corridor following, and floor changes. Tasks are grounded in MP3D and HM3D scene geometry, pairing short instructions with waypoint trajectories.

Evaluation

Benchmark results

The learned navigation tool

Fewer tool calls across both backends

For Claude Sonnet 5, the VLN tool improves success on both agentic benchmarks. For GPT-6 Astra, the success trade-off varies by task, while mean tool calls per episode decrease on both benchmarks.

Claude Sonnet 559.2% / 62.5%fewer calls · R2R / RxR

GPT-6 Astra46.9% / 51.0%fewer calls · R2R / RxR

Evolution results

The coding agent updates notes and executable skills over repeated passes through the same tasks. Model weights remain fixed.

R2R success72→98%

with VLN · pass 1 → 10

RxR success65→89%

with VLN · pass 1 → 10

Retained experienceNotes + code

fixed model weights

100 recurring tasks per benchmark, with retained experience and replay. Cost includes worker and improver charges.

BibTeX

@article{cheng2026asena,
  title   = {{ASENA}: Self-evolving Agents for Embodied Navigation},
  author  = {An-Chieh Cheng and Isabella Liu and Edmund Bu and Johan Bjorck and
             Hongxu Yin and Zhengyi Luo and Jan Kautz and Linxi {``Jim''} Fan and
             Yuke Zhu and Sifei Liu},
  year    = {2026},
  journal = {arXiv preprint arXiv:2609.39207},
  url     = {https://arxiv.org/abs/2609.39207}
}

Acknowledgements

We thank Jarred Travers, Amanpreet Singh, Peter Pham, Jinhyung (David) Park, Xiangchen Tian, and Tingwu Wang for their help with real robot infrastructure.

ASENA: Self-evolving agents for embodied navigation

Abstract

Real-world demonstrations

Skill creation and reuse