How SpaceX Propulsively Lands Rockets: The Science of Navigation, Guidance and Control

An in-depth engineering exploration of how SpaceX lands Falcon 9 boosters propulsively — covering hypersonic retropropulsion, the three-burn landing…

·

How SpaceX Propulsively Lands Rockets: The Science of Navigation, Guidance and Control

On December 21, 2015, a fifteen-story aluminum-lithium cylinder traveling at nearly 2 kilometers per second fell out of space, fired its engines into a hypersonic headwind, deployed four titanium waffle-irons as steering surfaces, and touched down — vertically — on a concrete pad at Cape Canaveral. The Falcon 9 booster's velocity at that moment was precisely zero meters per second. Its tilt was within a single degree of vertical. Its position was within a few meters of the painted X on Landing Zone 1. The engineering required to make this happen is arguably the most significant achievement in launch vehicle design since the Saturn V.

At the time of writing (July 2026), SpaceX has landed Falcon 9 boosters over 350 times. The feat that once seemed like science fiction is now a weekly occurrence. But the science that makes it possible — the interplay of hypersonic aerodynamics, real-time convex optimization, radar-based navigation, and millisecond-precision engine throttling — remains one of the most elegant engineering problems ever solved.

This guide goes deep: we'll trace every phase of the descent, unpack the physics at each step, and explain the navigation, guidance, and control (GNC) architecture that converts a ballistic missile trajectory into a gentle touchdown. No hand-waving. No "and then the rocket lands." We're going into the control loops.

1. Why Landing a Rocket Is Hard: The Problem Statement

Before we get to solutions, let's appreciate the problem.

At stage separation — roughly 2 minutes and 30 seconds after liftoff — the Falcon 9 first stage is:

The booster is now a supersonic, nearly-empty aluminum tube with nine rocket engines at one end, falling through the upper atmosphere at hypersonic speed — flying backwards. It has no wings. It has no parachutes. Its aerodynamic stability is, at best, marginal. Its center of pressure is in entirely the wrong place for forward flight, which is precisely what makes rear-first flight possible but also incredibly difficult to control.

Atmospheric entry, from the booster's perspective, involves:

  1. Flipping the vehicle 180° so the engines point into the velocity vector
  2. Bleeding off ~4,500 km/h of horizontal velocity
  3. Steering the vehicle toward a landing target hundreds of kilometers away (or exactly back at the launch site)
  4. Surviving reentry heating that would melt unprotected aluminum
  5. Reducing velocity to exactly zero at exactly the right altitude — with a vehicle that can't hover (its thrust-to-weight ratio > 1 even at minimum throttle)
  6. Landing on a 52 m × 91 m drone ship bobbing in the North Atlantic, or on a concrete pad precisely 91 meters in diameter

Each of these sub-problems requires a different branch of physics and control theory to solve. Let's take them in order.

2. The Three-Burn Landing Sequence

The landing sequence is choreographed around three precisely timed engine burns. Each uses a different number of engines and serves a distinct purpose.

Burn 1: Boostback Burn (3 engines, ~50 seconds)

Purpose: Reverse the downrange velocity component to bring the booster back toward the landing site. Only used for Return to Launch Site (RTLS) landings. Drone ship landings skip this burn — the booster continues ballistically toward the ship.

Physics: The booster is at ~80 km altitude, traveling ~6,000 km/h downrange. Three Merlin 1D engines reignite, producing ~2,550 kN of thrust. The engines fire into the velocity vector — this is literally "flying against your own momentum."

Delta-V imparted: ~1,500–2,000 m/s (varies by mission profile). This doesn't just zero out horizontal velocity — it actually reverses it, sending the booster back toward Florida. After the burn, the stage is on a parabolic trajectory that peaks at ~140–180 km before falling back toward the launch site.

A fascinating consequence: RTLS boosters essentially fly farther than orbital boosters. An RTLS booster goes up, turns around, comes halfway back, then descends — a longer total path than the expendable first stage that simply falls into the ocean.

Why 3 engines? Trading fuel efficiency for gravity losses. With only one engine, the burn would take too long, and the booster would lose significant altitude during the burn (requiring more fuel later to arrest the descent). Three engines minimize gravity losses while leaving enough propellant for the subsequent burns.

Burn 2: Reentry Burn (3 engines, ~20 seconds)

Purpose: Slow the booster from hypersonic velocity (~Mach 5–7) to supersonic (~Mach 2–3) before it hits the dense atmosphere. Without this burn, the booster would disintegrate — unlike a capsule with a heat shield, the aluminum fuselage can only survive reentry heating if the engines create a protective shockwave ahead of the vehicle.

Physics of Hypersonic Retropropulsion: This is the most counterintuitive part of the entire landing sequence. When the booster fires its engines INTO the hypersonic freestream, it creates a bow shock — a detached shockwave standing in front of the vehicle. This shockwave:

The rocket exhaust plume acts as a virtual aerospike — it pushes the shockwave further forward, reducing the heat flux reaching the vehicle body by 60–80%. SpaceX tested this explicitly during early recovery attempts: boosters without a reentry burn disintegrated. Boosters with it survived.

The physics here is a complex interaction between:

CFD simulations of this regime are notoriously difficult — it's one of the few flight regimes where computational predictions are still less reliable than flight test data. SpaceX effectively built a hypersonic wind tunnel that flies itself.

Burn 3: Landing Burn (1 engine, ~30 seconds)

Purpose: Final deceleration from terminal velocity (~300 m/s at 2–3 km altitude) to exactly 0 m/s at exactly 0 m altitude.

The Suicide Burn Problem: This is the famous "hoverslam" or "suicide burn" — so named because if you start the burn too late, you hit the ground at high speed; if you start too early, you come to a stop above the ground and then fall (since the engine can't throttle low enough to hover).

The governing equation is the one-dimensional rocket descent dynamics:

With the mass depleting as propellant is consumed:

Where \ddot{h} is the vertical acceleration (second derivative of altitude), T(t) is the time-varying thrust, m(t) is the instantaneous vehicle mass, g is gravitational acceleration (9.807 m/s² at sea level), m_0 is the mass at ignition, and \dot{m} is the propellant mass flow rate. The thrust itself is related to mass flow through T = \dot{m} v_e where v_e is the exhaust velocity.

The Falcon 9 at landing has a thrust-to-weight ratio of approximately 1.3–1.5 at minimum throttle. This means:

The minimum throttle constraint fundamentally determines the landing strategy. If TWR at minimum throttle were less than 1, the vehicle could simply descend slowly and hover just above the pad before touching down — like the Apollo Lunar Module. But because T_{min} / (m g) > 1, the engine must be ignited at precisely the right altitude so that the upward acceleration brings velocity to zero exactly at the surface. Start too early: the vehicle reaches v = 0 above the ground, then the engine — still producing more thrust than weight — accelerates it upward. You cannot throttle down enough to stop this. Start too late: lithobraking.

For constant deceleration (a reasonable approximation when the mass change during the short landing burn is modest), the ignition altitude is:

Where a = \frac{T}{m} - g is the net deceleration (typically 20–30 m/s², or 2–3 G). With initial velocity v_0 \approx 300 m/s at ~2 km altitude, this gives an ignition altitude of approximately 1,500–2,000 meters. The engine fires for roughly 30 seconds.

A more accurate expression for the burn time, accounting for mass depletion during the burn, comes from integrating the rocket equation over the landing maneuver:

Where v_e = I_{sp} \cdot g_0 \approx 2,750 m/s for the Merlin 1D (assuming I_{sp} \approx 280 s at sea level). For v_0 = 300 m/s, m_0 \approx 25,000 kg, and \dot{m} \approx 120 kg/s (single engine at ~40% throttle), this gives t_{burn} \approx 28 seconds — consistent with observed landing burns. The booster touches down at approximately 2–3 m/s; the landing legs absorb the residual energy.

The Merlin 1D engine throttles to approximately 40% of full thrust during landing (from a nominal 845 kN sea-level thrust per engine down to ~340 kN), achieved through a carefully designed pintle injector that maintains combustion stability across a wide throttling range.

3. Navigation: How the Booster Knows Where It Is

Accurate navigation is the foundation of the entire landing sequence. If you don't know where you are, you can't guide yourself anywhere. SpaceX uses a multi-sensor navigation suite:

GPS / INS Fusion

The primary navigation sensor is a tightly-coupled GPS + Inertial Navigation System (INS) :

The state vector being estimated is typically 15+ dimensions:

x = [position (3), velocity (3), attitude quaternion (4), 
     accelerometer bias (3), gyroscope bias (3), ...]

Radar Altimeter

In the final approach phase (below ~2 km), the booster switches from GPS altitude to a radar altimeter. This is critical because:

The radar is typically a Frequency-Modulated Continuous Wave (FMCW) design operating in the C-band or Ku-band, with a range accuracy of ~2–5 cm.

The Drone Ship Problem

When landing on an autonomous spaceport drone ship (ASDS) — "Of Course I Still Love You", "Just Read the Instructions", "A Shortfall of Gravitas" — the landing target is:

The booster must compensate for this motion in real-time. The ship transmits its GPS position and inertial state to the booster via a radio link, and the booster's guidance system treats the target as a moving reference frame — the landing burn aims for a predicted intercept point, not the current ship position.

The landing accuracy required: the Falcon 9's landing leg span is approximately 18 meters. The painted target circle on the drone ship is roughly 15 meters in diameter. On most landings, the booster touches down within 3–5 meters of center.

4. Guidance: How the Booster Decides Where to Go

Navigation tells you where you are. Guidance tells you where you need to go. Between them sits the most computationally impressive part of the landing system: the trajectory optimization that runs in real-time during flight.

The Powered Descent Guidance Problem

Formally, the problem is:

Find a thrust vector history T(t) that transfers the vehicle from its current state (r_0, v_0, m_0) to a target state (r_f, v_f) at touchdown, while minimizing fuel consumption and satisfying constraints on thrust magnitude, thrust direction, and glide-slope.

Mathematically, this is a non-convex optimal control problem with:

The use of \ln m (log mass) rather than m linearizes the mass depletion dynamics, which is essential for the convexification trick.

- Thrust magnitude bounded: T_{min} \leq \|\mathbf{T}(t)\|_2 \leq T_{max} - Thrust direction constrained: the vehicle must stay within a cone oriented opposite to the velocity vector (no flying sideways at Mach 3) - Glide-slope constraint (approach from above, not below the pad):

Where \theta_{gs} is the maximum glide slope angle (typically 20–30° from vertical) and \|\mathbf{r}(t)\|_2 = \sqrt{r_x^2 + r_y^2} is the horizontal distance from the target. - Final state: position error < 1 m, velocity < 0.5 m/s, attitude within 1° of vertical

For most of aerospace history, this problem was considered intractable for onboard real-time computation. Solutions required:

  1. Pre-computing a reference trajectory on the ground (Apollo LM approach)
  2. Perturbing from the reference using linear feedback (Space Shuttle approach)
  3. Solving the full problem but with simplified dynamics (early Mars lander approach)

SpaceX's innovation was realizing that through a mathematical technique called lossless convexification, the non-convex problem could be transformed into a convex one — and solved in milliseconds onboard.

Convex Optimization and CVXGEN

The breakthrough came from research at JPL and Stanford published around 2007–2010. The key insight: while the thrust magnitude constraint T_{min} \leq \|\mathbf{T}\|_2 \leq T_{max} is non-convex (it creates an annular feasible region — a ring — which is not convex), it can be relaxed by introducing a slack variable and proving that the optimal solution of the relaxed problem always lies on the boundary of the feasible set.

Simplified: instead of constraining the thrust to be exactly between T_{min} and T_{max}, you allow any thrust between 0 and T_{max}, then mathematically prove that fuel-optimal solutions will never use thrust below T_{min} anyway (because gravity loss is penalized). This transforms the problem from a difficult "mixed-integer" problem into a straightforward convex quadratic program.

The resulting optimization is a Second-Order Cone Program (SOCP) with the structure:

minimize    c^T x
subject to  ‖A_i x + b_i‖₂ ≤ c_i^T x + d_i,  i = 1,...,m
            F x = g

Where the decision vector x includes discretized positions, velocities, and thrust vectors at N timesteps into the future. In discretized form, the trajectory is represented as a sequence of N nodes with time step \Delta t between them:

Where \mathbf{x}_k = [r_x, r_y, r_z, v_x, v_y, v_z, \ln m]_k is the state at the k-th timestep and \mathbf{T}_k is the thrust vector applied during that timestep. The dynamics are enforced as linear equality constraints between consecutive timesteps, using the log-mass formulation that keeps the problem convex.

SpaceX uses a custom solver called CVXGEN (developed by Stanford's Stephen Boyd's group) that generates C code for exactly this problem structure. CVXGEN solves the SOCP in < 50 milliseconds on flight-qualified processors — fast enough to recompute the entire trajectory from scratch multiple times per second.

Convex Optimization in Action

The guidance system runs on a tight loop:

  1. Sample navigation state (position, velocity, orientation, mass) from the EKF
  2. Formulate the SOCP with current state as initial condition
  3. Solve via CVXGEN (~20–50 ms)
  4. Extract the first time-step thrust vector from the optimal solution
  5. Command the control system to achieve that thrust
  6. Shift the time horizon forward one step
  7. Repeat from step 1

This "receding horizon" approach means the booster is continuously replanning its entire trajectory, 10–20 times per second. If a gust of wind pushes the vehicle sideways, the next guidance solution compensates immediately. If a GPS measurement arrives with an updated position, the trajectory is recalculated around the new estimate.

This is fundamentally different from how Apollo landed on the Moon. Apollo used a pre-computed reference trajectory with linear-quadratic regulator (LQR) feedback — essentially "fly this path, and correct small errors." SpaceX's approach is "solve the complete optimization problem from scratch, dozens of times per second, using the actual current state." It's far more robust to large disturbances and initial condition uncertainty.

The Three Landing Modes

The guidance system handles three distinct mission profiles:

Profile · Downrange Distance · Boostback? · Landing Site · Fuel Usage

RTLS (Return to Launch Site) · 0 km · Yes (full) · LZ-1 / LZ-4 (concrete pad) · Highest

ASDS (Drone Ship) · 300–650 km · No · Drone ship at sea · Medium

Expendable · Unlimited · No · Ocean (no recovery) · None

For RTLS, the boostback burn targets a specific trajectory that brings the booster back to the launch site — the guidance system must coordinate all three burns. For ASDS, the booster continues ballistically after separation, and only the reentry and landing burns are active — but the target is moving, which adds complexity.

5. Control: How the Booster Actually Steers

Guidance produces a desired thrust vector. Control makes the vehicle actually produce that thrust — a fundamentally harder problem than it sounds.

Actuator Suite

The Falcon 9 booster has three independent control mechanisms during descent:

1. Engine Gimbaling (±5–7°)

The center engine (Engine 5 of 9 in the octaweb configuration) is mounted on a hydraulic gimbal that can tilt the engine nozzle by approximately ±5–7° in two axes. This provides:

The gimbal actuators are electro-hydraulic and respond at ~10–20 Hz bandwidth.

2. Cold Gas Thrusters (Nitrogen RCS)

Mounted near the top of the booster (and on the interstage), the Reaction Control System uses compressed nitrogen gas expelled through small thrusters to provide:

RCS thrust is modest (~400–900 N per thruster) but positioned far from the center of mass, giving strong torque authority.

3. Grid Fins (4 × Titanium)

The most visually distinctive control surfaces on the Falcon 9. Four lattice fins (nicknamed "grid fins" or colloquially "waffle irons") are mounted in an X-configuration at the top of the booster. They deploy after the reentry burn.

Why grid fins? Traditional planar fins would generate enormous drag and heating at hypersonic speeds. Grid fins:

The grid fins control pitch, yaw, and roll through differential deflection — rotating the fins in coordinated pairs generates moments about all three axes.

Material evolution: Early Falcon 9 flights used aluminum grid fins, which ablated and deformed during reentry. SpaceX switched to solid titanium castings — a single-piece titanium grid fin is one of the largest titanium castings in the world. Later iterations added thermal protection coatings.

Aerodynamic effectiveness: the grid fins produce ~0.5–1.0 G of lateral acceleration at supersonic speeds — enough to steer the booster precisely. Combined with engine gimbaling at low speeds, they provide full 6-DOF control authority throughout the descent.

Control Allocation

The control system solves a control allocation problem: given a desired force/torque vector from the guidance law, how do you distribute the effort across engine gimbal, RCS, and grid fins?

This is typically handled by solving a constrained quadratic program (smaller than the guidance SOCP):

minimize    ‖B u - τ_desired‖² + λ‖u‖²
subject to  u_min ≤ u ≤ u_max

Where B is the control effectiveness matrix (mapping actuator positions to forces/torques), u is the actuator command vector, and τ_desired is the desired torque from the guidance system.

The blend between aerodynamic surfaces and engine gimbaling shifts dramatically with airspeed:

Regime · Altitude · Primary Control · Secondary

Exo-atmospheric (>80 km) · Space · RCS thrusters · None

Hypersonic (80–40 km) · Upper atmosphere · Grid fins (weak) + RCS · Engine gimbal (during burns)

Supersonic (40–10 km) · Mid atmosphere · Grid fins (strong) · RCS

Transonic (10–2 km) · Lower atmosphere · Grid fins · Engine gimbal

Terminal (< 2 km) · Landing · Engine gimbal · Grid fins (fading)

The transition between regimes is managed by a blending function based on dynamic pressure q = ½ρv², ensuring smooth, bump-free control handoff.

6. Starship: The Next Generation

While Falcon 9 lands the first stage only, SpaceX's Starship system is designed for full reusability — both the Super Heavy booster AND the Starship upper stage must land propulsively.

Super Heavy Booster Landing: The "Chopstick" Catch

Rather than landing on legs (like Falcon 9), the Super Heavy booster is designed to be caught by the launch tower:

  1. The booster descends through its landing burn
  2. Two mechanical arms ("chopsticks") on the launch tower swing into position
  3. The booster positions itself between the arms, using small grid-like protrusions near the top that catch on the arms
  4. The arms close, supporting the booster's weight
  5. The booster is lowered directly onto the launch mount — no crane needed

This eliminates landing legs entirely (saving ~10% of dry mass) and allows the booster to be restacked on the launch mount within hours for the next flight. The catch mechanism requires positioning accuracy of mere centimeters — far tighter than Falcon 9's meter-scale tolerance.

The control problem is significantly harder: the booster must position itself not just over a flat pad, but within a gap between two moving mechanical arms, while compensating for wind, engine thrust asymmetry, and the tower's own structural flexure.

Starship Upper Stage: Belly-Flop and Flip

The Starship upper stage lands using an entirely different aerodynamic regime:

  1. Belly-flop descent: Starship falls horizontally through the atmosphere, presenting its broad windward side to generate maximum drag. This reduces terminal velocity from a ballistic ~3,000 m/s to a manageable ~70 m/s (terminal velocity in belly-flop configuration). This is a direct application of the rocket equation's corollary: if you can trade velocity for drag instead of propellant, you save enormous amounts of fuel.
  1. The flip maneuver: At approximately 500–1,000 meters altitude, Starship reignites two or three Raptor engines and executes a rapid 90° pitch rotation from horizontal to vertical. The engines gimbal aggressively to arrest the horizontal velocity while transitioning to vertical descent. The flip takes less than 3 seconds — the vehicle rotates at roughly 30°/second.
  1. Terminal landing: Identical to Falcon 9 — a single-engine or two-engine landing burn, with thrust vectoring for final positioning.

The belly-flop maneuver is unique to Starship. No other vehicle has ever attempted atmospheric entry in a high-drag horizontal orientation, transitioned to vertical, and landed propulsively. The control challenge is immense: during the flip, the vehicle's angle of attack sweeps from 90° to 0° in seconds, the aerodynamic center of pressure shifts violently, and the engines must simultaneously kill horizontal velocity and start building vertical thrust — all with sloshing cryogenic propellants.

SpaceX flight tested this extensively with Starship prototypes SN8 through SN15, achieving the first successful landing on SN15 in May 2021. The iterative approach — fly, crash, learn, fix, fly again — is exactly how they solved Falcon 9 landing a decade earlier.

7. Why This Matters: The Economics

The engineering is breathtaking, but the motivation is ruthlessly practical: money.

A Falcon 9 launch costs approximately 67 million (2026 pricing). The first stage accounts for roughly 60% of the vehicle's manufacturing cost — approximately 30–35 million. If you throw it away after every flight, that's $30+ million added to every launch.

By recovering and reusing the first stage, SpaceX:

The R&D program that achieved this cost SpaceX approximately $1 billion (2011–2017). They recouped that investment within approximately 2–3 years of operational reusability, and every landing since has been pure savings compared to the expendable alternative.

Starship aims to push this further: fully reusable, rapid turnaround (hours, not weeks), with payload capacity exceeding the Saturn V — at a per-launch cost target below $10 million. If achieved, it will reduce the cost of space access by another order of magnitude.

8. Key Takeaways

  1. Propulsive landing is a GNC problem, not a propulsion problem. The Merlin engine was capable of landing from the beginning. What took a decade of development was the navigation sensors, the guidance algorithms, and the control systems.
  1. Convex optimization is the secret weapon. The ability to solve the powered descent guidance problem in milliseconds onboard, using CVXGEN-generated SOCP solvers, is what distinguishes SpaceX's approach from every previous landing system.
  1. Hypersonic retropropulsion — using engine exhaust as a virtual heat shield — was an unproven technology before SpaceX. It is now the only viable method for landing large orbital-class boosters.
  1. Grid fins + engine gimbaling + RCS provide full 6-DOF control across the entire descent — the allocation between them shifts smoothly with dynamic pressure.
  1. The drone ship problem requires continuous replanning. A moving target that's rolling and heaving in 3-meter swells can't be captured with a pre-computed trajectory — the guidance system recalculates the entire landing, 10–20 times per second, right through touchdown.
  1. Starship extends this to full reusability with the belly-flop descent, the chopstick catch, and the flip maneuver — all harder versions of the Falcon 9 problem.

The next time you watch a Falcon 9 landing — that impossibly smooth transition from hypersonic fireball to gentle touchdown — you're watching the real-time solution of a seven-dimensional optimal control problem, computed in less time than it took you to read this sentence.

References: Lars Blackmore (SpaceX), "Autonomous Precision Landing of Space Rockets" (2016 Bridge article); Behcet Acikmese & Lars Blackmore, "Lossless Convexification of a Class of Optimal Control Problems with Non-Convex Control Constraints" (Automatica, 2011); John Carson III et al., "Powered Descent Guidance for Mars Landing" (JPL/Caltech, 2008); SpaceX webcast telemetry data; public FCC filings for drone ship communications.

Last updated: July 2026

More FabFlow blog posts