An LLM-enabled Multi-Agent Autonomous Mechatronics Design Framework¶
🕒 Published (v1): 2025-04-20 16:57 UTC · Source: Arxiv · Venue: CVPR 2025 · link
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
💬 Ask ChatGPT✦ Ask Claude
TL;DR¶
This paper proposes a hierarchical multi-agent framework that autonomously designs physical mechatronic systems by coordinating five specialized LLM-driven agents (planning, mechanical, simulation, electronics, embedded software) with structured human-in-the-loop feedback. The system is validated end-to-end by building a functional autonomous water-quality monitoring vessel. It defines a five-level "AMD Level" taxonomy analogous to SAE autonomy levels for engineering design.
Problem¶
Existing LLM-based agent frameworks operate in digital/simulated environments on narrow domains. Complex engineering tasks—mechatronics in particular—require cross-disciplinary integration (mechanics, electronics, control, firmware), constraint-aware reasoning about physical embodiment, and CAD/simulation tool interaction, none of which prior multi-agent systems address.
Method¶
A hierarchical orchestration architecture where a High-Level Planning Agent decomposes requirements into modular tasks via \(P = f(F, C, H)\) (functional requirements, constraints, human feedback), then dispatches to four domain agents:
- Mechanical Design Agent: LLM-driven parametric geometry generation using closed-form blade chord \(C(z)\) and pitch \(\alpha(z)\) distributions; generates CAD code via \(L: P \to C\) with iterative refinement \(P_{n+1} = F(P_n, R_n)\) from simulation feedback.
- Simulation & Validation Agent: Autonomously configures COMSOL CFD/FEA; runs laminar FSI baseline (von Mises stress \(\sigma_{vM}\)) then transitions to rotating-frame RANS with SST turbulence model for near-wall accuracy.
- Electronics Design Agent: Evaluates available hardware for reuse before designing new circuits; outputs PWM/H-bridge architecture for dual-motor control.
- Embedded Software Agent: Generates firmware from Electronics Agent specs; implements differential velocity steering without a physical rudder.
Human feedback gates ambiguous design choices at each stage rather than being a post-hoc review.
Key Contributions¶
- First LLM-enabled multi-agent framework targeting full mechatronics design across mechanical, electronics, control, and software domains simultaneously.
- AMD Level taxonomy (0–4) as a formalism for measuring design autonomy in physical engineering systems.
- Hybrid human-in-the-loop architecture that integrates structured feedback directly into agent planning loops rather than treating it as error correction.
- Demonstrated fabrication of a functional autonomous vessel from LLM-generated designs with CFD-validated propulsion.
Results¶
- Mechanical Design Agent iteratively converged propeller geometry from an initial poor-performing 3-blade design (26 mm blade, 20 mm hub) to a hydrodynamically optimized form improving lift-to-drag ratio and eliminating flow separation.
- Simulation & Validation Agent autonomously completed laminar FSI analysis; turbulent SST analysis required substantial human intervention for solver/boundary condition convergence.
- Embedded Software Agent produced firmware verified via logic analyzer: duty cycles of 39.2% (forward), 31.4% (backward), 58.8% (left/CCW), 15.7% (right/CW) at ~976 Hz, with only initialization-latency transients on first PWM pulse.
- No quantitative efficiency or time-savings baselines are reported; no ablation against human-designed counterpart.
Limitations¶
- No quantitative comparison to human-expert designs or prior automated pipelines; validation is qualitative/demonstrative.
- Simulation & Validation Agent cannot autonomously configure complex multiphysics (transient, turbulent) simulations—requires significant human expert correction.
- Mechanical agent lacks geometric/spatial reasoning; CAD output quality depends heavily on parametric formalism and iterative human feedback.
- Electronics and embedded agents validated only on a simple dual-motor platform; scalability to richer hardware stacks undemonstrated.
- Waterproofing, long-term deployment reliability, and hardware robustness are deferred to future work.
Relevance to Harnesses / Meta-Harnesses¶
This framework is a concrete domain-specific meta-harness: a fixed orchestration layer that routes subtasks to specialist agents, collects their outputs, and feeds results into downstream agents—exactly the pattern of a harness coordinating sub-processes. The AMD Level taxonomy is a formalization of autonomy that could inform how meta-harnesses decide when to invoke human feedback versus proceed autonomously. The hybrid human-in-the-loop gating mechanism (structured feedback injected at known bottlenecks rather than at the end) is a design pattern directly applicable to meta-harness architectures managing unreliable or capability-limited subagents.