🔬 applied physics

Saturation-Aware Robust Trajectory Optimization for Reusable Launch Vehicles via Differentiable Physics

This paper proposes a differentiable physics framework utilizing a Differentiable Particle Tube Control (DPTC) scheme to achieve robust trajectory optimization for reusable launch vehicles by embedding actuator saturation constraints directly into the computational graph, thereby outperforming conventional methods in maintaining closed-loop stability under aerodynamic uncertainties and control limits.

Original authors: Liwei Chen, Tong Qin

Published 2026-07-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Liwei Chen, Tong Qin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to land a giant, reusable rocket ship on a tiny pad. But here's the catch: the rocket has to do a wild, high-speed flip in the sky, like a gymnast doing a backflip while falling through a hurricane. The wind is gusting, the air is pushing the rocket in weird ways, and the engines have strict limits on how hard they can push or how far they can tilt.

The big question is: How do you program the rocket to land safely when everything is chaotic and the engines might hit their "max" limits?

This paper tackles that problem by comparing two different ways of thinking about the rocket's path.

The Old Way: The "Perfect World" Calculator

The first method, which the authors call CS-AD-SCvx, is like a super-smart calculator that assumes the rocket is in a perfect, predictable world. It draws a single, ideal line for the rocket to follow and then tries to keep the rocket on that line using a feedback loop.

The problem? This calculator is "saturation blind." It calculates how much the engines need to push to correct the rocket's path, but it forgets that the engines have a hard limit. It's like a GPS telling you to turn the steering wheel 90 degrees to stay on the road, not realizing your car's steering wheel can only turn 30 degrees. When the wind blows the rocket off course, this method demands more from the engines than they can physically give. The engines hit their limit, the math breaks, and the rocket starts to drift wildly off course.

In the paper's simulations, this "perfect world" approach led to a landing error where the rocket could miss the target by about 15.38 meters (roughly the length of a school bus) when the wind blew hard. The rocket's path became lopsided and unstable because the computer didn't account for the fact that the engines would get "stuck" at their maximum tilt.

The New Way: The "Crowd of Ghosts" Strategy

The authors propose a new method called Differentiable Particle Tube Control (DPTC). Instead of tracking just one perfect line, this method imagines the rocket as a cloud of thousands of "ghost" rockets (particles) all flying at once.

Think of it like this: Instead of trying to force one specific ghost to stay on a tightrope, the DPTC method shapes the entire cloud of ghosts. It asks, "How can we move the whole group so that even if some ghosts get pushed by the wind, they all stay within a safe zone?"

Crucially, this method is "saturation-aware." It knows the engines have limits. It learns that if the wind is too strong, it's better to let the rocket drift a little bit sideways (relaxing the path) rather than screaming at the engines to tilt further than they can go. By letting the cloud expand a bit in safe directions, it saves the engine's "muscle" for the most critical moments, like the final soft landing.

The Results: A Massive Improvement

When the authors ran 5,000 computer simulations to test these two methods against strong, unpredictable winds:

The old "perfect world" method failed to keep the rocket stable, resulting in a messy, wide landing area.
The new "cloud of ghosts" method (DPTC) kept the landing tight and symmetrical.

The paper reports that the new method reduced the landing error (specifically the 50% Circular Error Probable, or CEP50) by 87%. Instead of missing the target by 15.38 meters, the new method landed within just 1.97 meters of the center.

Why This Matters

The paper shows that for extreme maneuvers like this "belly-flop" flip, you can't just rely on linear math that ignores physical limits. By using a "differentiable physics" engine—essentially a computer simulation that can calculate exactly how small changes affect the whole flight—the new method learns to make smart trade-offs. It sacrifices a tiny bit of horizontal precision to ensure the rocket doesn't run out of control authority and crash.

The authors note that while this new method takes more computing power to train (running thousands of simulations offline), the final result is a simple, fast set of instructions that a rocket can use in real-time. It's a shift from trying to force a rocket to follow a rigid line, to teaching it how to dance with the wind while respecting its own physical limits.

Technical Summary: Saturation-Aware Robust Trajectory Optimization for Reusable Launch Vehicles via Differentiable Physics

1. Problem Statement

The vertical powered landing (VPL) of reusable launch vehicles (RLVs), particularly the high-angle-of-attack "flip maneuver" (e.g., the Starship architecture), presents a critical challenge for robust guidance, navigation, and control (GNC). This phase involves highly nonlinear aerodynamic coupling, large structural loads, and rapid dynamic pressure variations. The primary difficulty lies in ensuring landing precision and vehicle survivability under severe multi-source uncertainties (unsteady aerodynamics, wind shear, mass uncertainties) while adhering to strict physical actuator limits (thrust vector control gimbal bounds and engine throttling limits).

Existing robust trajectory optimization methods, such as Successive Convexification (SCvx) combined with Covariance Steering (CS), face two fundamental limitations in this context:

Local Linearization Assumptions: They rely on Gaussian covariance propagation derived from local linearization, which fails to accurately capture the non-Gaussian, skewed state distributions induced by severe nonlinear aerodynamic transients.
Saturation Blindness: They typically synthesize feedback laws (e.g., Riccati-based LQR) under the assumption of infinite control authority. When disturbances drive required control commands beyond physical actuator bounds, the resulting hardware truncation invalidates the theoretical robustness guarantees, often leading to actuator saturation, loss of control authority, and degraded landing precision.

2. Methodology

The paper proposes a Differentiable Particle Tube Control (DPTC) framework, which integrates differentiable physics with ensemble-based optimization to address these limitations. The methodology is built upon a native Differentiable Physics Engine that couples 6-DoF rigid-body kinematics with a deep-learned neural aerodynamic surrogate.

Core Components:

Differentiable Physics Engine: The system models 6-DoF flight dynamics using a 4th-order Runge-Kutta integrator where all operations (including quaternion normalization and neural network forward passes) are formulated as differentiable tensor operations. This allows for exact gradient propagation through the entire nonlinear dynamic graph.
Stochastic Distribution Shaping: Unlike traditional methods that optimize a single nominal trajectory, DPTC treats robust guidance as a constrained stochastic probability transport problem. It represents the uncertain state distribution using a Lagrangian particle ensemble ( $N_p$ particles). Each particle evolves independently under unique disturbance realizations but shares a unified control policy.
Saturation-Aware Feedback Synthesis: The control policy is parameterized as a time-varying affine feedback law ( $u = u_{ref} - K(x - x_{ref})$ ). Crucially, hard actuator projection operators ( $\Pi_U$ ) are embedded directly into the forward computational graph. This forces the optimizer to synthesize the feedforward trajectory ( $u_{ref}$ ) and feedback gains ( $K$ ) simultaneously while "experiencing" the same actuator saturations that occur in the real vehicle.
Tail-Risk Optimization: To handle safety without the brittleness of hard chance-constraints, a differentiable tail-risk objective is introduced. Using a smooth surrogate (e.g., $ReLU^2$ ), the framework penalizes particles violating safety constraints while maintaining a continuous gradient landscape for backpropagation through time (BPTT).

Comparison Baseline:

The proposed method is benchmarked against an Automatic Differentiation-based Successive Convexification (CS-AD-SCvx) baseline. This baseline uses local linearization to extract Jacobians, propagates Gaussian covariance via Riccati equations, and tightens constraints using chance-constraint margins, but it lacks native saturation awareness during feedback synthesis.

3. Key Contributions

The paper outlines three primary contributions:

Transition from Trajectory Optimization to Distribution Shaping: The framework moves beyond classical Gaussian-linearization approximations to operate directly on exact nonlinear dynamics. By leveraging a natively differentiable 6-DoF engine, it shapes the stochastic probability manifold, inherently capturing rare-event dispersions and non-Gaussian aerodynamic responses.
Native Saturation-Aware Feedback Synthesis: Unlike covariance steering methods that append feedback post-hoc, DPTC embeds hard physical actuator limits directly into the optimization graph. The optimizer synthesizes the control policy while respecting the same saturation constraints that govern the physical vehicle, preventing the generation of infeasible control commands.
Soft-Penalty Tail Risk Regulation: The introduction of a differentiable tail-risk objective allows for smooth gradient propagation through time. This enables the optimizer to aggressively penalize constraint-violating particles (tail risk) without the numerical instability associated with hard chance-constraints, yielding a highly survivable closed-loop guidance law.

4. Results

The framework was evaluated using 6-DoF Monte Carlo simulations (5,000 runs) under 5% unsteady aerodynamic disturbances.

Nominal Performance: In deterministic conditions (zero disturbance), both DPTC and CS-AD-SCvx achieved nearly identical fuel-optimal trajectories, with final vehicle mass differences of less than 0.25%. This validated the fidelity of the differentiable physics engine.
Robustness under Uncertainty:
- CS-AD-SCvx: Under disturbances, the baseline exhibited significant performance degradation. Its unconstrained feedback formulation led to actuator saturation (violating deep-throttling and gimbal limits). This caused a mismatch between the assumed linear covariance evolution and the physically executed dynamics, resulting in asymmetric landing dispersion and a 50% Circular Error Probable (CEP50) of 15.38 m.
- DPTC: The proposed framework proactively performed a constraint-aware performance trade-off. By relaxing spatial tracking requirements during high-dynamic-pressure phases, it preserved critical control authority for the terminal descent. This resulted in a symmetric envelopment of the nominal trajectory and a CEP50 of 1.97 m, representing an 87% reduction in landing error compared to the baseline.
Computational Efficiency: While DPTC requires significant offline computation (GPU-accelerated parallel rollout of particle ensembles), the resulting policy is a time-varying tensor sequence. Online deployment reduces to $O(1)$ lookup and matrix multiplication, making it suitable for real-time execution. The framework demonstrated $O(1)$ scaling with respect to batch size on GPUs due to vectorization.

5. Significance and Claims

The paper claims that integrating differentiable physics with ensemble-based optimization provides an effective and practical framework for robust guidance in highly constrained aerospace flight systems.

Paradigm Shift: The work demonstrates a transition from deterministic trajectory tracking to physics-informed stochastic distribution shaping. It argues that under severe uncertainties and strict actuator limits, the optimal strategy is not to rigidly track a path, but to actively shape the stochastic state distribution through nonlinear dynamics.
Practical Robustness: By embedding physical constraints directly into the learning process, the framework prevents the "saturation blindness" inherent in classical robust control methods. This ensures that the guidance law remains valid and survivable even when the vehicle operates at the boundaries of its physical envelope.
Scalability: The approach offers a scalable architecture for high-dimensional, nonlinear aerospace problems, utilizing exact analytical gradients derived from the physics engine rather than relying on idealized linear approximations or finite-difference methods.

The authors conclude that this framework significantly enhances mission robustness and physical constraint satisfaction, offering a viable alternative for advanced GNC systems in next-generation reusable launch vehicles.