Controlled Agentic Planning & Reasoning for Mechanism Synthesis

TL;DR

This paper tackles automated planar mechanism design by coupling dual-agent LLM-driven linguistics with simulation-based evaluation in a closed-loop framework. The proposed approach formalizes a design-critique cycle where a design agent generates executable mechanism hypotheses, a critic agent evaluates them against physics-based simulations and memory, and symbolic regression provides interpretable trajectory surrogates to guide refinement. The authors introduce MSynth, a benchmark of analytically defined planar trajectories, and demonstrate that critic feedback yields substantial improvements (up to 90% distance reduction) while architecture, scale, and memory exhibit model-specific effects. The work advances neuro-symbolic planning for mechanism synthesis and provides a scalable, memory-augmented framework for linguistically grounded optimization with potential implications for automated engineering design.

Abstract

This work presents a dual-agent \ac{llm}-based reasoning framework for automated planar mechanism synthesis that tightly couples linguistic specification with symbolic representation and simulation. From a natural-language task description, the system composes symbolic constraints and equations, generates and parametrises simulation code, and iteratively refines designs via critic-driven feedback, including symbolic regression and geometric distance metrics, closing an actionable linguistic/symbolic optimisation loop. To evaluate the approach, we introduce MSynth, a benchmark of analytically defined planar trajectories. Empirically, critic feedback and iterative refinement yield large improvements (up to 90\% on individual tasks) and statistically significant gains per the Wilcoxon signed-rank test. Symbolic-regression prompts provide deeper mechanistic insight primarily when paired with larger models or architectures with appropriate inductive biases (e.g., LRM).

Figures (17)

• Figure 1: Example windshield-wiper synthesis workflow. Engineers first generate a candidate mechanism, use these to build simulations during a dimensional-synthesis step, and then iteratively refine topology and dimensions based on simulation performance until the design meets targets.
• Figure 2: llm-guided pipeline for synthesising planar mechanisms with a quarter-circle end-effector path: the da generates proposals using domain knowledge and memory; the simulator returns numeric metrics and a sr trajectory; the ca provides recommendations; simulator+critique feedback loop back to the DA for iterative refinement until a valid design emerges.
• Figure 3: Average percentage improvement in Chamfer distance when comparing runs with feedback (Fdbk) versus without feedback. Incorporating feedback yields substantial gains, with the largest uplift observed for llamas (from 30% to 46%), demonstrating that feedback consistently enhances performance.
• Figure 4: Neuro-symbolic closed-loop for planar mechanism synthesis: (1) da generates executable mechanism hypotheses from the specification and memory; (2) Executor runs candidates and ca critiques and issues structured revision directives; (3) da performs the revision, then repeats. The circle in the picture represents the path traced by the mechanism as simulated by the simulator.
• Figure 5: Ground truth Ellipse: the original target Ellipse used for evaluation.