HexaShell coding-agent harness sets a new RoboDojo score
Original title: HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents
With Codex and GPT-6-Astra, success is 39.7% on 42 tasks; tight insertion is 8.3% while pushing reaches 92.5%.
HexaShell is a thin harness that lets a vision-language coding agent drive a two-arm robot from one command line, using three calibrated RGB cameras and code tools it writes after failed attempts. Tools are built on ten layouts of one seed per task, then frozen and tested on ten new layouts. Using Codex with GPT-6-Astra at medium reasoning effort, the authors report a new RoboDojo state of the art and say failures are mostly contact, not perception or reasoning. Tools evolved so far are open sourced.
Unconfirmed: The state-of-the-art claim is the authors' leaderboard comparison, not an independent audit.
- Reported progress score is 48.0, averaged like the leaderboard.
- Pushing reaches 92.5% and memory tasks 65.0%; tight insertion is 8.3%.
- Tools use triangulation and plane-ray intersection from robot observations only.
- AI models
- GPT-6-Astra
From the original article
1 min read at huggingface.co
Agents built on vision language models (VLMs) now compete with learned policies on robot manipulation leaderboards. We ask how far a coding agent can go when it controls a robot instead of a computer, and where it stops. Our system, HexaShell, keeps the harness thin: it only turns the agent's commands into robot motion.
By Hongcheng Gao, Alex Zhou, Jay Zhu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.