Embodied agents improve robot navigation and reasoning with coding and vision
ASENA: Self-evolving Agents for Embodied Navigation
Robotics
Summary
Robots need to move around and do tasks by themselves, but this is hard because they must understand their surroundings and follow instructions. The authors created ASENA, a system that lets robots write and fix their own programs using what they see and remember, without changing the underlying AI model. They added a special navigation policy called ASENA-VLN that helps the robot predict where to go by looking at images and language instructions. This system works better than previous approaches at navigating and answering questions in both simulated and real environments.
What this means in practice
- •For robot developers: Use ASENA to enable robots to program and improve their navigation and interaction skills autonomously using vision and language feedback.
- •For mobile robot operators: Deploy ASENA-powered robots to perform complex navigation, spatial reasoning, and task execution in dynamic real-world spaces without pre-built maps.$Commercial implications: ASENA enables autonomous robots with advanced self-improving navigation and multi-tasking abilities suitable for commercial service and inspection robots.
Authors
An-Chieh Cheng, Isabella Liu, Edmund Bu, Johan Bjorck, Hongxu Yin, Zhengyi Luo, Jan Kautz, Linxi "Jim" Fan, Yuke Zhu, Sifei Liu
Abstract
We present ASENA, an embodied agent system that connects general-purpose coding agents to robot sensing, computation, supervised execution, and persistent experience. Agents can write and execute programs, inspect recorded outcomes, repair failures, and reuse notes and executable skills while keeping their model weights fixed. We further introduce ASENA-VLN, a 4B monocular navigation policy that serves as an optional tool within this programmable system. ASENA-VLN predicts body-frame trajectories for both extended routes and short-horizon behaviors using a shared vision-language decoder trained on route instructions, visual question answering, and a newly curated dataset of geometry-derived atomic navigation tasks. As a standalone policy, ASENA-VLN achieves state-of-the-art success rates of 68.7% on R2R and 70.2% on RxR. When integrated with a coding agent, learned navigation improves ASENA's success rate by 11 percentage points on both agentic benchmarks while reducing execution time. Through persistent workspace evolution and simulator feedback, ten passes over recurring 100-task subsets further improve success from 72% to 98% on R2R and from 65% to 89% on RxR. On embodied question answering, ASENA achieves state-of-the-art accuracy with fewer interaction steps. Finally, real-world demonstrations on a Unitree G1 combine search, visual inspection, spatial reasoning, and synthesized gestures without a pre-built map, illustrating how online programming extends robot behavior beyond route following and predefined skills.