Research map / Reinforcement learning / Visual and representation RL
NeuralPlane: An Efficiently Parallelizable Platform for Fixed-wing Aircraft Control with Reinforcement Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces NeuralPlane, a GPU-accelerated fixed-wing aircraft simulation platform (MLP-approximated aerodynamic tables) for large-scale parallel RL training, with PPO baselines and traditional Ardupilot-based controllers.
Problem
RL for fixed-wing aircraft control needs many training samples, but existing high-fidelity flight simulators are CPU-bound and do not scale to large parallel simulation.
Approach
A GPU-accelerated flight dynamics model that stores states and aerodynamic parameters as tensors and replaces aerodynamic lookup tables with MLP approximations, enabling simulation of up to 10^6 aircraft in parallel, bundled with task scenarios, baselines, and evaluation/visualization tools.
Key results
- Single-step simulation of 10 s takes about 0.2 seconds at 10^6 parallel aircraft, versus other platforms (JSBSim, ArduPilot, X-Plane) in Figure 1.
- MLP fits of aerodynamic coefficients reach test R2 of 0.9926 to 0.9991 across Cl, Cm, Cn, Cx, Cz.
- PPO achieves over 95% success rate across Normal, HighSpeed, HighAltitude, Windy, and Noisy Heading scenarios and outperforms the traditional controller on maneuverability metrics, though the traditional controller is slightly better on safety metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPU-accelerated FDM for large-scale parallel simulation | NeuralPlane · Parallel simulation with 10^6 aircraft | single-step simulation time: 0.2 seconds | Helped |
| Increase the number of parallel rollouts | PPO training in NeuralPlane · Heading task | training convergence | Helped |
| Use PPO instead of the traditional controller | F16 control · Heading task under Normal, HighSpeed, HighAltitude, Windy, and Noisy conditions | maneuverability | Helped |
| Use the traditional controller instead of PPO | F16 control · Heading task under tested conditions | safety metrics | Helped |
| Fit aerodynamic parameters with an MLP in place of table lookups | NeuralPlane FDM · Aerodynamic parameter fitting | test-set R2: 0.9949 ± 0.0009; 0.9962 ± 0.0007; 0.9926 ± 0.0011; 0.9967 ± 0.0004; 0.9991 ± 0.0001 | Helped |
| Increase observation noise to scale 0.01 | PPO · Robustness test | time to reach target: 9.52s; 9.09s; 9.52s; 9.52s; 11.11s | Hurt |
Surprising findings
- With only 32 parallel rollouts PPO fails to converge and reward fluctuates strongly, while 3000 parallel rollouts converge in about 100 iterations.
- The traditional controller slightly outperforms PPO on safety metrics even though PPO is better on maneuverability.
Limitations
- Supports a limited number of aircraft models and task scenarios.
- No multi-aircraft scenario support yet.
- User-friendliness and interfaces still need improvement.
Experimental setup
Benchmarks: NeuralPlane Heading/Control/Tracking task scenarios
Metrics: G TAS RoC AOA time to complete task altitude/speed/overload/AOA/sideslip safety margins success rate time to reach target under noise
Baselines: JSBSim ArduPilot X-Plane TECS + L1 PID controller PPO
Builds on: PPO JSBSim QPlane Isaac Gym
Closest papers by meaning
- A deep inverse-mapping model for a flapping robotic wing (ICLR 2025 · similarity 0.84)
- UAV-Flow Colosseo: A Real-World Benchmark for Flying-on-a-Word UAV Imitation Learning (NeurIPS 2025 · similarity 0.84)
- WFCRL: A Multi-Agent Reinforcement Learning Benchmark for Wind Farm Control (NeurIPS 2024 · similarity 0.83)
- SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores (ICLR 2024 · similarity 0.82)
- Improving Autonomous Separation Assurance through Distributed Reinforcement Learning with Attention Networks (AAAI 2024 · similarity 0.82)
- Adaptive Surrogate Gradients for Sequential Reinforcement Learning in Spiking Neural Networks (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.