Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLMs for Engineering: Teaching Models to Design High Powered Rockets

Toby Simonds

TL;DR

This work investigates applying language-model–based systems to high-powered rocket design using RocketBench, a bridge to the RocketPy 6-DOF simulator. It defines two design tasks—target apogee optimization and precision landing—and shows baseline LLMs possess substantial engineering knowledge but struggle to iteratively improve designs from simulation feedback, with performance plateauing below human experts. By training with Group Relative Policy Optimization (GRPO) on a 7B Qwen model, the authors achieve superhuman performance, attaining precision landings within 12 m and peak scores exceeding human baselines on both tasks. The results suggest RL-augmented LLMs can serve as powerful optimization tools in engineering domains when interfaced with domain-specific simulators, enabling accelerated design cycles and novel capabilities beyond traditional software-focused tasks.

Abstract

Large Language Models (LLMs) have transformed software engineering, but their application to physical engineering domains remains underexplored. This paper evaluates LLMs' capabilities in high-powered rocketry design through RocketBench, a benchmark connecting LLMs to high-fidelity rocket simulations. We test models on two increasingly complex design tasks: target altitude optimization and precision landing challenges. Our findings reveal that while state-of-the-art LLMs demonstrate strong baseline engineering knowledge, they struggle to iterate on their designs when given simulation results and ultimately plateau below human performance levels. However, when enhanced with reinforcement learning (RL), we show that a 7B parameter model outperforms both SoTA foundation models and human experts. This research demonstrates that RL-trained LLMs can serve as effective tools for complex engineering optimization, potentially transforming engineering domains beyond software development.

LLMs for Engineering: Teaching Models to Design High Powered Rockets

TL;DR

This work investigates applying language-model–based systems to high-powered rocket design using RocketBench, a bridge to the RocketPy 6-DOF simulator. It defines two design tasks—target apogee optimization and precision landing—and shows baseline LLMs possess substantial engineering knowledge but struggle to iteratively improve designs from simulation feedback, with performance plateauing below human experts. By training with Group Relative Policy Optimization (GRPO) on a 7B Qwen model, the authors achieve superhuman performance, attaining precision landings within 12 m and peak scores exceeding human baselines on both tasks. The results suggest RL-augmented LLMs can serve as powerful optimization tools in engineering domains when interfaced with domain-specific simulators, enabling accelerated design cycles and novel capabilities beyond traditional software-focused tasks.

Abstract

Large Language Models (LLMs) have transformed software engineering, but their application to physical engineering domains remains underexplored. This paper evaluates LLMs' capabilities in high-powered rocketry design through RocketBench, a benchmark connecting LLMs to high-fidelity rocket simulations. We test models on two increasingly complex design tasks: target altitude optimization and precision landing challenges. Our findings reveal that while state-of-the-art LLMs demonstrate strong baseline engineering knowledge, they struggle to iterate on their designs when given simulation results and ultimately plateau below human performance levels. However, when enhanced with reinforcement learning (RL), we show that a 7B parameter model outperforms both SoTA foundation models and human experts. This research demonstrates that RL-trained LLMs can serve as effective tools for complex engineering optimization, potentially transforming engineering domains beyond software development.
Paper Structure (27 sections, 2 equations, 5 figures, 1 table)

This paper contains 27 sections, 2 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Key Contributions
  3. Related Work
  4. RL for Engineering Design
  5. Methodology
  6. Simulation Env
  7. Competition Tasks
  8. Target Altitude Challenge
  9. Precision Landing Challenge
  10. RL Dynamics
  11. Iterative Prompting Protocol
  12. Human Baseline
  13. Results
  14. Rocket Bench
  15. Target Altitude Challenge
  16. ...and 12 more sections

Figures (5)

  • Figure 1: Performance across Target Altitude Challenge. Left shows average performance across 10 runs right shows sample progress over time (Best over 3 runs for LLMs)
  • Figure 2: Performance comparison of Claude vs. o1 over 30 iterations with best human performance indicated.
  • Figure 3: Performance across Precision Landing Challenge. Left shows average and best performance across 10 runs right shows sample progress over time (Best over 3 runs for LLMs)
  • Figure 4: Comparison of RL performance. Reward represents mean batch reward
  • Figure 5: Best Rocket Design from RL on Precision Landing Challenge