Back to feed
News Story
APriority81
机器之心
1 sources

Embodied Navigation Large Models Trained in Vain? Coding-Agent Barely Connected to Robot Achieves 78% Success Rate, Surpassing Industrial-Grade Specialized Models

A study by Qi Wu's team at the University of Adelaide directly connected coding agents (such as Claude Code and Codex) to robots without mapping, memory, or training, achieving a 78% success rate on the R2R-CE benchmark for vision-and-language navigation, surpassing industrial-grade specialized models. The research challenges traditional navigation methods, suggesting that general-purpose reasoning models can directly control embodied agents.

SynthePulse Insight · AI deep readingMembers

Embodied Navigation LLMs Trained in Vain? Coding Agents Bare-Wired to Robots Hit 78% Success, Outperforming Industrial-Grade Specialized Models

Version 1 · 1 source

A new study from Qi Wu's team at the University of Adelaide directly connects coding agents to robots—no mapping, no memory, zero training—achieving 78% success on the R2R-CE benchmark, matching or even surpassing industrial-grade specialized models. This result challenges the assumption in embodied navigation that specialized training or complex scaffolding is necessary, but long-horizon navigation remains a blind spot.

  • On 100 random paths in R2R-CE val-unseen, zero-shot coding agents achieve 78% success, compared to 94% for humans.
  • The model determines performance, not the framework: switching models can swing results by over 20 points, while switching harnesses only changes a few points.
  • Forcing an external waypoint predictor can backfire: strong models lose points on tasks where they originally had high SR; but when offered as an optional tool, fable-5 rises from 68.3% to 76.7%, with half the steps and a quarter of the time.
Open section navigationCounterintuitive Result: Bare-Bones Coding Agent Matches Industrial-Grade Models

Counterintuitive Result: Bare-Bones Coding Agent Matches Industrial-Grade Models

A new study from Qi Wu's team at the University of Adelaide takes coding agents (like Claude Code, Codex) and connects them to robots as-is, providing only a monocular camera and four action buttons, with no mapping, no memory, and no navigation training. On the classic visual language navigation (VLN) benchmark R2R-CE, they achieve 78% success, matching or even surpassing recent industrial-grade navigation models trained on large-scale data; for reference, humans score 94% on this benchmark.

This result challenges the common assumption of two orthodox approaches in the field: that general-purpose large models can't navigate on their own, so either you train them (e.g., NaVid, StreamVLN, NavFoM) or you scaffold them (e.g., NavGPT, MapGPT, SmartWay). The proposed MIP (Minimal-Interface Probe) method only keeps two functions, observe() and step(), letting the reasoning model decide each step itself, which the paper calls agentic embodied control.

In a real replay of episode 10 from R2R-CE, the agent bumps into a bed corner, then notices 'I've barely moved,' and proactively adjusts its angle to go around, demonstrating autonomous error correction.

Free for now

Read the full analysis

4 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

This article's information primarily comes from a Machine Intelligence report on the paper, which is a secondhand account. Specific data (such as 78%, 84%) are from the paper or the report and have not been independently verified. Some conclusions (like 'scaffolding value depreciates') are inferences by the research team, and the report does not provide independent verification.

Sources for this version

  1. 具身导航大模型白训了？coding-agent裸接机器人，成功率78%反超工业级专用模型机器之心

Primary report

机器之心

Primary source