High-Altitude Balloon Station-Keeping with First Order Model Predictive Control

Myles Pasetsky, Jiawei Lin, Bradley Guo, Sarah DeanarXiv 2025 · DOI · arXiv

Introduction

High-altitude balloons (HABs) are cost-effective platforms for atmospheric research and space observation. However, maintaining a balloon's position over a specific target—a task known as station-keeping—is exceptionally challenging. HABs are underactuated, capable only of vertical movement to harness varying wind currents at different altitudes. Furthermore, they operate in partially observable environments where wind forecasts are frequently inaccurate.

Conceptual diagram of balloon station-keeping showing the discrepancy between predicted and actual trajectories and the role of replanning.

Historically, the complexity of balloon dynamics and wind uncertainty led researchers toward model-free Reinforcement Learning (RL). While RL has demonstrated success, it often lacks a rigorous model-based baseline. This paper introduces First-Order Model Predictive Control (FOMPC), demonstrating that a differentiable model-based approach can significantly outperform state-of-the-art RL agents.

Methodology: First-Order Model Predictive Control

The core of FOMPC is a differentiable simulator implemented in JAX, encompassing balloon dynamics, wind models, and cost functions. This differentiability allows the controller to optimize flight paths in real-time using first-order gradient descent.

The FOMPC architecture, featuring heuristic-based initialization and first-order optimization within a receding horizon framework.

Key technical innovations include:

  • Differentiable Dynamics: Models ranging from simple drift to complex thermal and battery dynamics are implemented to support automatic differentiation.
  • Continuous Action Relaxation: FOMPC plans using continuous actions, which are subsequently mapped to the physical constraints of discrete venting and pumping.
  • Heuristic-based Initialization: To navigate non-convex optimization landscapes and avoid local minima, the planner initializes with multiple "smart guesses," such as "coast" baselines and random altitude-seeking trajectories.
  • Receding Horizon Planning: The agent plans several hours ahead but only executes the initial actions before replanning with updated wind data.

Experimental Setup

The researchers evaluated FOMPC using the Balloon Learning Environment (BLE), an open-source simulator.

  • Wind Data: The environment utilizes the ERA5 reanalysis dataset for historical global climate data.
  • Uncertainty Modeling: Simplex noise is added to forecast winds to create a "ground truth" distinct from the controller's predictions.
  • Benchmarking: The system was tested over 10,000 randomized scenarios and compared against Google’s Perciatelli44 RL agent.

Results and Key Insights

FOMPC achieved a 39.43% Time-Within-Radius (TWR), a 24% relative improvement over the 31.78% TWR of the Perciatelli44 agent. This improvement translates to approximately 1.8 additional hours per day within the target zone. Furthermore, FOMPC significantly enhanced safety, reducing power violations from 22.14 to 8.70 by explicitly penalizing low-battery states in the cost function.

Key insights from the study include:

  • Planning vs. Reacting: FOMPC’s look-ahead capability allows it to exploit future wind patterns that reactive RL agents might overlook.
  • The Power of Replanning: Frequent replanning (e.g., every 72 minutes) effectively compensates for simplified or inaccurate wind models.
  • Model Fidelity: Surprisingly, even lower-fidelity dynamics models outperformed RL, suggesting that trajectory optimization, rather than perfect physics modeling, is the primary performance bottleneck.

Discussion and Future Work

While FOMPC is computationally more intensive than RL (~0.13s vs ~0.001s per step), it remains well within the real-time requirements for balloon control. However, gradient-based optimization can still struggle with highly complex wind fields.

Future research may explore hybrid approaches, such as using FOMPC as a "teacher" for Imitation Learning to train RL agents. This could combine the strategic foresight of MPC with the rapid execution and robustness of RL policies.

Reference: Pasetsky, M., Lin, J., Guo, B., & Dean, S. (2025). High-Altitude Balloon Station-Keeping with First Order Model Predictive Control. arXiv:2511.07761v1 [cs.RO].

Summary updated

BibTeX
@misc{pasetsky2025high,
  author = {Myles Pasetsky and Jiawei Lin and Bradley Guo and Sarah Dean},
  title = {{High-Altitude Balloon Station-Keeping with First Order Model Predictive Control}},
  year = {2025},
  eprint = {2511.07761},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.07761},
}
Figure