@article{shi2026maestro,
  title = {{MAESTRO}: Orchestrating Robotics Modules with Coding Agents for Zero-Shot Generalist Robots},
  author = {Junyao Shi* and Rujia Yang* and Kaitian Chao* and Bingqing Selina Wan and Yifei Simon Shao and Jiahui Lei and Jianing Qian and Long Le and Pratik Chaudhari and Kostas Daniilidis and Chuan Wen and Dinesh Jayaraman},
  year = {2026},
  month = {Nov},
  pub_type = {conference},
  abstract = {Today's dominant route to generalist robots scales up "observations-in, actions-out" teleoperation datasets to train large end-to-end models, echoing the recipe behind vision-language models (VLMs). We pursue a road less traveled: building generalist policies directly around VLMs, augmenting their broad capabilities with a curated set of perception, planning, and control modules. In MAESTRO, a VLM coding agent dynamically composes these modules into a programmatic, closed-loop policy that monitors execution and replans autonomously. It shares the coding-agent pattern of prior code-as-policies (CaP) systems, but unlike a coding agent acting on files it controls, a robotics agent controls neither its tools nor the open, dynamic world they act in, so any tool can fail unpredictably. MAESTRO's core contribution is a robotics-specific harness that makes composing curated external robotics models actually work, wrapping error-prone tool chains in robust calls, rendering their outputs back for the VLM to verify, and closing the loop to recover from failures. Ablations show that adding the same tools without this harness barely helps and can even drop performance below the original no-tool baseline. Across challenging tabletop and mobile manipulation tasks, and with no new teleoperation data, this lets a zero-shot coding agent match or surpass state-of-the-art vision-language-action (VLA) models, with the largest gains in settings underrepresented in VLA training data, making MAESTRO the first competitive modular generalist policy. MAESTRO can further call a VLA as one of its tools and autonomously generate real-world trajectories for downstream learning. We argue that building out this kind of agentic system is a complementary, far less-explored axis to data scaling for general-purpose robots. Project site: maestro-robot.github.io.},
  journal = {CORL},
  url_pdf = {/publication/shi-2026-maestro/shi-2026-maestro.pdf},
  url_project = {https://maestro-robot.github.io/},
  url = {https://maestro-robot.github.io/static/Maestro.pdf},
  url_arxiv = {https://maestro-robot.github.io/static/Maestro.pdf},
}