Deep Learning Warm Starts for Trajectory Optimization on the International Space Station

TL;DR

Onboard trajectory optimization for space systems is computationally demanding on resource-constrained flight computers. The authors employ amortized optimization to learn a problem–solution mapping that provides warm starts for Sequential Convex Programming (SCP), with safety ensured online by the GuSTO solver; the neural-warm-starts are trained offline from optimized trajectories. Key findings show that learned warm starts preserve the optimal cost $J^*$ while reducing solver iterations by about $60\%$ for rotational dynamics and about $50\%$ for obstacle scenarios drawn from the training data, though generalization to unseen obstacles is limited by the network architecture ($J^*$ remains unchanged but convergence accelerates only in familiar regimes). This work demonstrates safe, onboard, learning-guided trajectory optimization in space, elevating the Technology Readiness Level from $3$ to $5$ and enabling real-time autonomous guidance, navigation, and control for future missions.

Abstract

Trajectory optimization is a cornerstone of modern robot autonomy, enabling systems to compute trajectories and controls in real-time while respecting safety and physical constraints. However, it has seen limited usage in spaceflight applications due to its heavy computational demands that exceed the capability of most flight computers. In this work, we provide results on the first in-space demonstration of using machine learning-based warm starts for accelerating trajectory optimization for the Astrobee free-flying robot onboard the International Space Station (ISS). We formulate a data-driven optimal control approach that trains a neural network to learn the structure of the trajectory generation problem being solved using sequential convex programming (SCP). Onboard, this trained neural network predicts solutions for the trajectory generation problem and relies on using the SCP solver to enforce safety constraints for the system. Our trained network reduces the number of solver iterations required for convergence in cases including rotational dynamics by 60% and in cases with obstacles drawn from the training distribution of the warm start model by 50%. This work represents a significant milestone in the use of learning-based control for spaceflight applications and a stepping stone for future advances in the use of machine learning for autonomous guidance, navigation, & control.

Figures (5)

• Figure 2: Overview of our learning-based warm-starting approach for accelerating trajectory optimization. The baseline method uses Sequential Convex Programming (SCP) initialized with a cold start (i.e., a straight-line trajectory). In the offline phase, a neural network is trained on solved planning problems to predict effective warm starts. At deployment (online), the learned warm start is used to initialize SCP, resulting in faster convergence to an optimal trajectory.
• Figure 3: Testing phases for software on Astrobee. Left: Ground testing on a 2D air-bearing granite table at NASA Ames to validate software on real hardware in a controlled planar environment. Right: On-orbit flight testing aboard the ISS, where Astrobee executed autonomous trajectories in a 6-DOF microgravity environment.
• Figure 4: Iterations to convergence for cold and warm starts across trajectory categories. Horizontal bars indicate mean iterations required for planner convergence under cold and warm start conditions, with error bars showing one standard deviation. Warm starts perform similar to cold starts for translation-only trajectories. Warm starts require fewer iterations than cold starts for trajectories with rotation and seen obstacles, i.e., non-convexities learned by the model. However, the model did not generalize to unseen obstacles.
• Figure 5: Examples of trajectories tested. Top row shows Astrobee execution and bottom row shows the plotted data. Both cold and warm starts result in the same optimal trajectory. Warm starts are qualitatively closer to the optimal trajectory, especially for trajectories with obstacle avoidance and rotations.
• Figure 6: Comparison of convergence iterations for cold and warm starts. Each pair of connected points represents a matched run, i.e., a cold and warm start for the same start and goal states and obstacle constraints. The plot shows a trend towards reduction in iterations required for convergence when using a warm start. Black markers indicate group means, with vertical bars showing 95% confidence intervals computed using the standard error of the mean.