Qumus

Embodied multi-agent AI system that autonomously plans, fabricates, and nano-processes atomically thin two-dimensional quantum materials and van der Waals device stacks inside a robotic mini-laboratory.

   
Affiliation Princeton University Department of Physics and Princeton AI Lab (Sanfeng Wu group), with University of Michigan CSE, NIMS Tsukuba (Watanabe and Taniguchi), and California State University Northridge
First introduced 2026-05 (arXiv:2605.18407)
Lifecycle stages Multi-stage (hypothesis generation, protocol planning, multi-step robotic execution, optical analysis, reporting)
Autonomy level Fully autonomous — natural-language scientific goals trigger closed-loop fabrication and characterization with no human intervention beyond providing raw materials and electricity
Domain focus Quantum materials — graphene, hBN, transition-metal dichalcogenides, and van der Waals stacks; field-effect transistor fabrication
Availability Code on request — paper states code will be released on GitHub (link to be provided); demo videos at qumus.ai

Approach

Multi-agent system embodied within a custom robotic minilab. A lead Qumus agent orchestrates four specialized sub-agents: Project Manager (literature, experimental history, registered Skills and recipes), Lab Manager (material inventory, instrument status, computer-vision tool positioning), Device Expert (vdW stacking order, positions, twist angles), and Processing Agent (executable laboratory actions). The Processing Agent operates a three-tier hierarchical workflow: fixed “Atom Workflows” (robotic-arm control, stage movement, camera, vacuum, tape handling); composable “Molecule Workflows” (exfoliation, automated flake search, relocation); and “Assembly Workflows” (repeated exfoliation, stacking, longer fabrication sequences) — AI self-generates the molecule and assembly levels while atom-level actions remain user-fixed. The hardware platform integrates automated Scotch-tape exfoliation, optical inspection with multi-magnification microscopy, precision 2D crystal transfer and vdW stacking, and sample storage in a single workstation. YOLO-based macroscopic vision plus micro-QR fiducials and motorized-focus optical microscopy provide multi-scale perception. The system was characterized across six leading LLM backbones (GPT, Gemini, Claude, Grok, Qwen, DeepSeek).

Validation

End-to-end demonstrations on real hardware: AI-driven creation of graphene flakes, hexagonal boron nitride exfoliation, and AI fabrication of atomically thin field-effect transistors via vdW stacking. In an open-ended task (“I want a graphene flake larger than 200 μm²” with cleared history), Qumus (Claude Sonnet 4.6) explored the parameter space of substrate heating temperature, dwell time, massage cycles, and tape peel-off speed across five consecutive experimental runs over four hours, eventually producing a flake meeting the goal. The authors quantify each LLM’s experimental personality along seven dimensions (protocol alignment, caution, bias for action, token efficiency, agent efficiency, consistency, report quality).

Notable results

  • First reported AI creation of graphene and first AI fabrication of complex nanodevices including atomically thin field-effect transistors via vdW stacking.
  • Demonstrated autonomous error detection and recovery (e.g., recovering after a human removed an in-process chip mid-experiment, and after a Processing Agent hallucination mislabeled hBN as graphene).
  • Goal-oriented long-horizon execution with closed-loop hypothesis–experiment–analysis iteration over multi-hour timescales.

Primary paper

Shi et al., “Qumus: Realization of An Embodied AI Quantum Material Experimentalist,” arXiv:2605.18407.

Other references

None yet.

Code

GitHub release stated as forthcoming (link to be provided). Demo videos at qumus.ai.