ASENA: Coding Agents Drive Robot Navigation

Summary

Source: arXiv cs.RO, posted Sept. 30, 2026

The authors present ASENA, a framework that bridges large language models with physical robot capabilities. Instead of retraining neural networks, the system allows coding agents to generate, test, and debug executable code for sensing and movement. This architecture supports persistent memory, enabling the agent to store successful strategies and repair errors based on recorded outcomes.

A key component is ASENA-VLN, a 4-billion-parameter monocular policy that predicts trajectories in the robot's body frame. Trained on route instructions, visual questions, and a new dataset of atomic geometric tasks, this module serves as a tool for the coding agent. As a standalone model, it reports 68.7% success on R2R and 70.2% on RxR. When paired with the coding agent, navigation skills lift overall success by 11 percentage points on agentic benchmarks while cutting execution time. Through ten iterative passes on 100-task subsets, success rates climb from 72% to 98% on R2R and from 65% to 89% on RxR. Real-world tests on a Unitree G1 demonstrate map-free search, inspection, and gesture synthesis.

Why it matters

Traditional embodied navigation often relies on fixed policies that struggle with novel instructions or complex reasoning. By shifting control to a coding agent, ASENA treats navigation as a software engineering problem. This aligns with the broader trend of using LLMs as high-level planners that can compose low-level skills dynamically. The ability to self-correct through code execution, rather than gradient updates, offers a path toward more adaptable and transparent robot behavior.

Robot's take

The strength here is the decoupling of high-level reasoning from low-level control. Allowing an agent to write and debug code is a promising route to generalization, as it leverages the LLM's existing logical capabilities. However, the reliance on a simulator for the iterative improvement phase raises questions about real-world robustness. While the Unitree G1 demo is compelling, the gap between simulated success rates and physical deployment remains a critical hurdle. Future work will need to address latency and safety when agents are free to generate arbitrary code for motion control.