JARVIS: An Agent That Designs, Builds, and Tests Circuit Boards on Its Own

Shibo Zhou, Tianqi Hong, Zheng Yan, He YangarXiv 2026 · arXiv

TL;DR

JARVIS is an embodied AI agent that autonomously designs, builds, and tests circuit boards, closing the "design-build-test-learn" loop for the first time in power electronics. By integrating large language model (LLM) reasoning with a self-built digital twin and robotic fabrication, it can transform a plain-language request into a functional, verified hardware device in under three hours without any human intervention.

What Problem Does This Paper Address?

While software development has been accelerated by AI and automation, hardware development remains "stuck at the speed of human hands." An engineer must manually design a circuit, layout the board, operate various fabrication machines, wire up test equipment, and then interpret the results to start the next iteration.

Existing AI tools for electronics often stop at the design or simulation stage. The physical "loop"—the actual printing, assembly, and measurement—is a massive bottleneck. JARVIS addresses this by creating an agent that acts as both engineer and technician, handling everything from the initial requirement to the final performance analysis and subsequent redesign.

The JARVIS autonomous loop: from a plain-language need to a tested board and the next design iteration.

Methods

JARVIS utilizes a multi-layered architecture to bridge the gap between abstract reasoning and physical action:

  • The Agent: Powered by an LLM (Claude Opus 5.5), the agent handles high-level planning, code generation, and decision-making. It maintains a "long-term memory" through text-based records of every machine and task.
  • Executable Digital Twin: Before touching any materials, JARVIS uses a robot-mounted 3D scanner to survey its workspace and build a digital twin. This twin is "executable," meaning it runs the exact same code as the real machines, allowing the agent to "rehearse" every motion, thermal profile, and circuit response in simulation.
  • Accountable Execution Layer: This layer manages "machine skills"—the specific programs for tool changes, printing, and testing. It enforces safety limits and logs every action in an "attempt ledger," making the process auditable and resumable.
  • Physical Bench: JARVIS operates unmodified desktop equipment, including a 6-axis robot arm, a PCB printer (using conductive silver ink), a pick-and-place machine, and professional-grade Hardware-in-the-Loop (HIL) test equipment (OPAL-RT simulator and oscilloscope).
The physical JARVIS bench alongside its autonomously generated digital twin.

Datasets

The paper introduces a 60 Hz notch filter as a benchmark for evaluating the closed-loop system. This task was chosen because its requirements are precise and its success can be measured exactly. The agent was tasked with designing a filter to remove 60 Hz mains hum from a signal while maintaining a flat passband. The system's performance was also benchmarked against ten other representative "self-driving laboratory" systems across various scientific fields.

Key Insights and Ideas

  • Rehearsal is Essential: The "rehearsal in the twin" caught five major defects before any material was used. For example, it identified a test plan that would have incorrectly rejected a good board and a printing pattern that would have shorted the circuit.
  • Accountability via the Ledger: By treating every action as an "attempt" with a stated goal, plan, and expected result, the agent can autonomously recover from faults. If a machine stops responding, the agent reads the state, chooses a remedy, and continues.
  • Self-Surveying World Model: Rather than relying on hard-coded coordinates, the agent builds its own world model from 3D scans. This allows it to adapt to changes in the physical environment (like a moved machine or a slightly bent board) automatically.
  • Closing the Loop: JARVIS doesn't just provide a "pass/fail" verdict. It analyzes the measured data (e.g., a sag in low-frequency gain), identifies the physical cause (output impedance mismatch), and automatically generates the next design iteration in KiCad to fix the issue.
The architecture of JARVIS, showing how the agent interacts with the digital twin and the physical bench.

Key Results

In a full end-to-end run, JARVIS achieved the following:

  • Speed: Completed a 42-step design-build-test cycle in 2 hours and 39 minutes with zero human intervention.
  • Performance: The fabricated notch filter achieved 50.6 dB attenuation at 60 Hz, significantly exceeding the 40 dB requirement.
  • Autonomous Troubleshooting: During the run, the agent resolved two real-world machine faults (an unrecognized tool head and a frozen application) by reading machine states and applying remedies.
  • Iterative Learning: The agent correctly diagnosed why the first board's passband sagged at 1 Hz and proposed a "stiffer" circuit design for the next iteration, which it then drew in KiCad and rehearsed in the twin.

Limitations and Future Work

The authors acknowledge several areas for improvement:

  • Design Aesthetics: The agent currently uses 90-degree trace bends and non-optimal routing paths that a human engineer would avoid, though these did not affect electrical performance at 60 Hz.
  • Complexity: The current demonstration is limited to signal-level analog circuits. Future work will target more complex power stages and embedded controllers.
  • Generality: While the framework is general, specific "machine skills" and circuit models must be developed for new classes of hardware.

The ultimate goal is to evolve JARVIS into a professional-grade platform that can serve as a "self-driving laboratory" for the rapid development of the power electronics needed for the global energy transition.

Summary updated

Abstract

Power electronics is central to the energy transition, but developing new hardware still relies on human-led design, fabrication, testing, and iteration. We present JARVIS, an embodied AI agent that autonomously closes this loop for circuit boards. JARVIS maintains machine-specific capabilities and constraints, scans its workspace using a robot-mounted 3D scanner, and constructs an executable digital twin in which it rehearses robot motions, fabrication steps, and circuit behavior before consuming materials. A single text instruction initiates the physical experiment. Using unmodified desktop equipment, the agent coordinates a robot arm, a PCB printer that deposits conductive silver ink and solder paste, and a pick-and-place machine. It verifies progress using cameras and sensors, then tests the fabricated circuit with an OPAL-RT real-time simulator and an oscilloscope. Based on the measurements, JARVIS diagnoses deviations from the design and generates a revised design in KiCad for subsequent rehearsal. In a demonstration involving a 60 Hz notch filter for grid-connected converter sensing, JARVIS completed a first-run experiment in 2 h 39 min without human intervention, achieving 50.6 dB attenuation at 60 Hz. Its digital-twin rehearsal identified five issues before fabrication, including a test criterion that would have incorrectly rejected a functioning board. During execution, the agent also resolved encountered faults autonomously. These results demonstrate an integrated agent-driven workflow for physical design, fabrication, measurement, and iterative improvement of power-electronic circuitry. To our knowledge, JARVIS is the first system in power electronics in which an agent closes the physical design-build-test-learn loop for circuit boards.
BibTeX
@misc{zhou2026jarvis,
  author = {Shibo Zhou and Tianqi Hong and Zheng Yan and He Yang},
  title = {{JARVIS: An Agent That Designs, Builds, and Tests Circuit Boards on Its Own}},
  year = {2026},
  eprint = {2611.26503},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26503},
}
Figure