Skip to content

RocketSmith: Agentic Additive Manufacturing of High-Powered Rockets

🕒 Published (v1): 2026-06-26 00:00 UTC · Source: HuggingFace · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

RocketSmith is an agentic system (released as a Claude Code MCP plugin) that automates the design-for-additive-manufacturing (DFAM) pipeline for high-powered rockets — from flight simulation and stability analysis to parametric CAD generation and slicer-based mass estimation. Four rockets were built and flight-tested, all achieving stable launch and two recovered re-flyable, with altimeter data showing ~80% of predicted apogee.

Problem

High-powered rocket development requires iterative loops across flight simulation, CAD, slicing, and assembly, but there is significant friction between these software tools (OpenRocket, CAD packages, slicers). Manual iteration delays design convergence when manufacturing variance (actual mass, tolerance shifts) forces recalculation of stability and subsequent part redesign. Prior work (RocketBench) showed that LLMs plateau below human expert performance when asked to iteratively optimize design parameters directly from simulation feedback.

Method

RocketSmith operates as a Claude Code plugin using the Model Context Protocol (MCP) and is composed of 6 subagents: rocketsmith (orchestration), openrocket (simulation), cadsmith (CAD generation), prusaslicer (slicing/mass estimation), manufacturing (DFAM/DFM handoff), and gui (visualization).

The pipeline is: 1. OpenRocket builds a component tree from user constraints (motor, dimensions, flight profile); stability is calculated via the Barrowman equation. 2. build123d (parametric Python CAD library over OpenCASCADE) generates deterministic STEP files from the OpenRocket-derived dimensions. 3. PrusaSlicer estimates precise part mass from the sliced toolpath (material, infill, layer height) — these actual masses override OpenRocket's estimates and trigger re-verification of stability. 4. Parts are FDM-printed, assembled, and flight-tested.

A key architectural choice (motivated by RocketBench findings) is that numerically precise tasks — stability calculation, trajectory simulation, mass estimation — are delegated to deterministic tools rather than left to LLM reasoning. The LLM orchestrates but does not simulate.

Supports both zero-shot and human-in-the-loop workflows.

Key Contributions

  • First end-to-end agentic system spanning flight simulation → parametric CAD → slicing → manufacturing for high-powered rocketry.
  • Design principle validated: delegating precision tasks to tools (not LLM reasoning) closes the iteration loop without the performance plateau identified by RocketBench.
  • Open-sourced MCP-based plugin architecture (6 subagents, skills, GUI dashboard) reusable for other physical design domains.
  • Empirical flight-test validation (not just simulation metrics) of agent-designed hardware.

Results

  • 4 distinct high-powered rockets built, using AeroTech H100W, H219T (×2), and J425R motors.
  • All 4 achieved stable launch (no weathercocking, tumble, or CATO).
  • 2 of 4 recovered in re-flyable condition (recovery system failures on the other 2, not design or structural failures).
  • Altitude accuracy: Rockets reached ~80% of the apogee predicted by the agentic system, establishing consistency between simulation and experiment.
  • Components fabricated via FDM on consumer-grade printers, manually evaluated for flight readiness.

Limitations

  • 50% recovery rate — recovery deployment reliability not yet within the agent's optimization scope.
  • 20% gap between predicted and actual apogee (unmodeled drag, mass variance, atmospheric effects).
  • Limited to FDM (PLA/PETG/etc.); no support for advanced AM (SLS, SLA, metal) or composite layup.
  • Only Class 2 (H–J impulse) rockets tested; scalability to Class 3 or orbital-class vehicles unvalidated.
  • Manual assembly and post-processing still required — tolerance fit issues resolved by the human, not the agent.
  • GUI is read-only visualization; all interaction still goes through the CLI harness.

Relevance to Agentic AI / LLM Agents

RocketSmith is a concrete case study in physical-domain agentic systems: it demonstrates how tool-augmented LLM agents can bridge the gap between simulation and manufacturing in a safety-critical iterative design loop. Its key insight — that the LLM should orchestrate but not perform precision computation — echoes the broader agentic-AI principle of delegating verifiable subtasks to deterministic tools. The MCP-plugin architecture and subagent decomposition (orchestrator + domain specialists + GUI) offer a reusable pattern for any engineering DFAM pipeline.