RocketSmith: An Agentic System for High-Powered Rocket Design and Manufacturing

RocketSmith is an agentic system that automates the design, additive manufacturing, and optimization of high-powered rockets, successfully validating its capabilities through the fabrication and flight testing of four distinct vehicles that achieved stable launches and high simulation accuracy.

The Gist

Imagine you want to build a high-powered rocket, but instead of having a team of engineers, a CAD designer, a flight simulator expert, and a manufacturing specialist, you have just one very smart, very fast robot assistant. That is RocketSmith.

This paper describes a new "agentic system" (a fancy term for a team of AI robots working together) that can take a simple idea from a human, design a rocket, figure out how to build it, and even run simulations to see if it will fly safely—all before a single piece of plastic is printed.

Here is how RocketSmith works, broken down into everyday concepts:

1. The "Brain" and the "Specialists"

Think of RocketSmith not as one robot, but as a project manager who hires a team of specialized interns.

• The Project Manager (The Main Agent): This is the AI that listens to what you want (e.g., "I want a rocket that flies 1,000 feet high using this specific motor").
• The Specialists (Subagents): The manager doesn't do the math or draw the blueprints itself. It calls on specific sub-agents:
  • The Flight Simulator: It uses software called OpenRocket to pretend the rocket is flying. It checks if the rocket will wobble, how high it will go, and if it will land safely.
  • The Architect (CADSmith): Once the flight plan looks good, this agent draws the 3D blueprints using code. It turns the math into a digital 3D model.
  • The Factory Foreman (PrusaSlicer): This agent looks at the 3D model and figures out exactly how to print it on a 3D printer. It calculates how much the part will weigh and how much plastic is needed.

2. The "Loop" of Perfection

Building a rocket is tricky. If you print a part and it's slightly heavier than you thought, the rocket might become unstable.

• The Old Way: An engineer would design a rocket, print it, weigh it, realize it's too heavy, go back to the computer, change the design, print it again, and repeat this cycle for weeks.
• The RocketSmith Way: The system does this loop instantly. It designs a rocket, simulates the flight, prints a digital version of the part to weigh it, realizes "Oh, this is 5 grams too heavy," and immediately tweaks the design to fix it. It does this "zero-shot" (on its own) or with a human giving it a nudge ("Make the walls thicker").

3. The Real-World Test Drive

The researchers didn't just talk about the system; they built four actual rockets using RocketSmith and launched them.

• The Materials: They used 3D printers (like the ones hobbyists use) to print the rockets out of plastic (ABS and PETG).
• The Results:
  • All four rockets flew. They didn't fall apart immediately; they launched straight up.
  • Two rockets landed safely and were ready to fly again. One landed in a tree (but was still in one piece), and another landed softly in a field.
  • Two rockets had issues: One broke apart because a parachute opened too late (a timing error), and another crashed because the parachute didn't open at all.
  • The "Crystal Ball" Accuracy: For the rockets that had sensors, the AI's prediction of how high they would go was surprisingly accurate. One rocket went to 80% of the predicted height, and another hit 84%.

4. The "Human in the Loop"

The paper emphasizes that while the AI is powerful, it's not magic.

• Sometimes the AI needs a human to double-check things, especially when it comes to the very specific details of how to print the plastic so it doesn't crack.
• Think of it like a co-pilot. The AI does 90% of the heavy lifting (designing, calculating, planning), but the human pilot (the engineer) makes the final call on safety and handles the tricky physical assembly.

The Bottom Line

RocketSmith is a tool that turns the complex, messy, and time-consuming process of building high-powered rockets into a streamlined, automated workflow. It proved that an AI system can design a rocket, generate the instructions to build it, and create a vehicle that actually flies. While it didn't get everything perfect (two rockets crashed), it successfully turned a digital idea into a flying machine, showing that AI can be a powerful partner in engineering complex physical objects.

Technical Summary

Technical Summary: RocketSmith – An Agentic System for High-Powered Rocket Design and Manufacturing

Problem Statement

The design and manufacturing of launch vehicles are inherently multifaceted challenges requiring specialized domain knowledge in propulsion, mechanical structures, and materials. While traditional development relies on collaborative interdisciplinary teams, high-powered rocketry (Class 2) often scales these processes down to individual efforts. This reduction in scale introduces significant friction between the various software tools required for flight simulation, parametric design, and manufacturing. Specifically, the iterative loop between design specifications and manufacturing realities (e.g., part availability, actual component mass, and stability recalculations) is often prolonged by manual coordination and internal discussion. There is a need for a system that can streamline this iterative development cycle, automating the transition from flight simulation to parametric CAD generation and manufacturing file creation while maintaining stability and performance constraints.

Methodology

The authors present RocketSmith, an agentic system built upon the Model Context Protocol (MCP) and released as a plugin for agent harnesses like Claude Code and Gemini-CLI. The system orchestrates the design, simulation, and manufacturing workflow through a combination of subagents, skills, and tool-calling capabilities.

Software Architecture and Tool Integration

The core pipeline integrates three primary software tools:

1. OpenRocket: Serves as the "source of truth" for rocket design and flight simulation. The system utilizes the OpenRocket Helper package to programmatically access the Java-based platform, allowing the agent to create component trees, modify mass/dimensions, and run flight simulations to predict apogee, velocity, and stability.
2. build123d: A parametric CAD library wrapping OpenCASCADE. Unlike generative 3D approaches (e.g., diffusion models), RocketSmith uses build123d to generate deterministic, script-based STEP files. This allows for precise geometric adjustments (e.g., changing a fillet radius) without regenerating the entire solid and ensures the output is an industry-standard exchange format suitable for downstream slicing.
3. PrusaSlicer: Used to slice the generated STEP files for Fused Deposition Modeling (FDM). The system uses this tool to obtain precise weight estimations based on specific material (PETG or ABS), infill, and printer configurations, feeding this data back into OpenRocket for stability recalculations.

Agentic System Components

• Subagents: The system employs six specialized subagents (rocketsmith, cadsmith, gui, manufacturing, openrocket, praslicer) to compartmentalize context. This prevents context window pollution and allows specific agents to manage API interactions for their respective domains (e.g., the manufacturing agent handles Design for Additive Manufacturing (DFAM) checks).
• Skills: Seven distinct skills guide the workflow: motor-selection, stability-analysis, design-for-additive-manufacturing, generate-structures, modify-structures, mass-calibration, and print-preparation. The workflow typically begins with motor selection, followed by stability verification, structural generation, minor modifications (e.g., adding screw holes), and finally, preparation for slicing.
• Human-in-the-Loop: The system operates in both zero-shot and human-in-the-loop modes. A Graphical User Interface (GUI) visualizes the agent's state, component trees, and CAD models, allowing users to intervene, request adjustments (e.g., thicker walls), and validate outputs before manufacturing.

Experimental Setup and Manufacturing

The system was evaluated by designing and manufacturing four distinct high-powered rockets (High Power 1–4) with varying motor configurations (AeroTech H100W, H219T, and J425R) and assembly complexities.

• Materials: Components were fabricated using FDM printers (Ender 3, Ender 5 Plus, and Voron-2-Tall) with PETG and ABS filaments, selected for their heat deflection temperatures.
• Design Variations: The rockets ranged from simple two-part assemblies to complex multi-stage designs with avionics bays. High Power 2 was specifically designed to fit within the build volumes of commercial printers, while High Power 1 and 4 utilized large-format printing for longer airframes.
• Flight Testing: All four rockets were flight-tested at Dragon's Fire Field under the supervision of the Tripoli Rocketry Association. Two rockets (High Power 1 and 4) were equipped with StratoLogger CF altimeters and RunCam 5 cameras for telemetry and video recording.

Results

The flight tests yielded mixed but informative results regarding the system's capability to produce flight-ready hardware:

• Launch Stability: All four rockets achieved stable launches with no off-axis tipping or weathercocking during the boost phase.
• Recovery Outcomes:
  • High Power 3: Successfully recovered in reflyable condition after landing in a tree (split occurred during retrieval, not flight).
  • High Power 4: Successfully recovered in reflyable condition with minimal effort.
  • High Power 1: Suffered a catastrophic structural failure in the lower airframe during recovery deployment due to a delayed ejection charge, though electronics were recovered.
  • High Power 2: Failed to separate the recovery bay, resulting in a ballistic impact and structural damage.
• Flight Performance Metrics:
  • Apogee Accuracy: For the instrumented rockets, the measured apogee closely matched simulations. High Power 1 reached 80% of the expected altitude, and High Power 4 reached 84%.
  • Velocity and Acceleration: Measured values for High Power 4 were slightly lower than predicted (92.4 m/s vs. 115.2 m/s max velocity), likely due to aerodynamic drag or mass variations not fully captured in the initial simulation.

Key Contributions and Significance

The paper claims that RocketSmith demonstrates the feasibility of using agentic systems to autonomously manage the complex, iterative lifecycle of high-powered rocket development.

• Automation of Iteration: The system successfully closed the loop between simulation, CAD generation, and manufacturing weight estimation, reducing the friction typically associated with manual tool switching.
• Design for Manufacturing (DFM): By integrating PrusaSlicer weight data back into OpenRocket, the system allowed for iterative stability optimization based on realistic manufacturing constraints rather than theoretical values.
• Validation of Agentic Workflows: The successful launch and partial recovery of four distinct rockets validate that an LLM-based agent can handle the domain-specific reasoning required for mechanical design and assembly planning.

The authors conclude that while the system significantly reduces development time and friction, human oversight remains critical for specialized tasks such as finalizing complex electronics covers and ensuring manufacturing tolerances. The work establishes a foundation for future agentic systems in domain-specific engineering applications, proving that such systems can produce designs suitable for real-world manufacturing and flight testing.