Small-Scale Rocket Landing: Can a Learned Guidance Policy Outperform a Classical Approach in Simulation?

Powered descent guidance for a rocket, whether full-scale or small-scale, is a control problem in which a vehicle must reach a target position and velocity while burning as little propellant as possible, subject to physical limits. Two broad families of methods exist to solve it: classical trajectory optimization, most often convex optimization, and learned policies trained with reinforcement learning (RL). The research question asks whether a trained neural network guidance policy can match or beat a classic convex-optimization law such as lossless convexification (LCvx) on a small-scale reusable rocket, in landing accuracy and fuel efficiency.

The claims collected here describe RL guidance systems and convex-optimization guidance systems, but none of them puts the two side by side on the same vehicle, same scenario set, and same accuracy or fuel metric. Some claims describe RL performance on a 6-DOF planetary lander [1], others describe timing and constraint-satisfaction gains for learned warm-starts to convex solvers [4][11], and others describe convex optimization used alone [14][10]. These are different systems, tested under different conditions, so their numbers cannot be placed on a common scale.

This matters for a builder because the two families solve the guidance problem differently. Convex optimization methods, when applicable, give guarantees: deterministic behaviour, global optimality, and certificates of convergence or infeasibility [14]. RL policies do not carry these guarantees by construction, but they are reported to produce accurate and fuel-efficient trajectories in simulation, and to be robust to noise and parameter uncertainty [1]. Whether that robustness and accuracy amounts to comparable or better performance than a specific convex law like LCvx, on a specific small vehicle, is exactly the missing head-to-head test.

The scale of the vehicle also matters, and none of the claims are stated for a 10 kg, 100 m apogee rocket specifically. The RL results below come from a 6-DOF planetary lander [1], from a high-fidelity rocket model [6], from lunar and asteroid scenarios [3][8]. The convex-optimization results come from Mars powered landing formulations [14][4] and from a rocket powered landing problem with an angle-of-attack constraint [10]. Applying any of these numbers to a small-scale vehicle without re-testing would be an extrapolation the sources do not support.

The rest of this note lays out what each family of method actually does, what has been measured for each, and what a two-person team would need to build to answer the original question directly, since the literature does not yet answer it.

An RL guidance policy for rocket landing is a neural network trained to output thrust commands directly from the vehicle's estimated state. In one approach, Proximal Policy Optimization (PPO) is used to learn a policy that maps the lander's estimated state directly to a commanded thrust for each engine [1]. Training used a 6-DOF simulation environment for the descent and landing task, and the resulting policy produced trajectories described as accurate and fuel-efficient [1]. The system also showed robustness to noise and to system parameter uncertainty [1].

A specific tuning detail reported for this system is that using different discount rates for terminal and shaping rewards significantly enhances optimization performance [1]. This is stated as a general enhancement to the training procedure, tied to the reward-shaping design used in that particular 6-DOF lander study, rather than as an isolated one-off result. A builder adopting a similar reward structure, with both a terminal reward for reaching the target and a shaping reward that guides intermediate behaviour, should expect to need to tune these two discount rates separately rather than sharing one value between them.

Other RL guidance work targets related but distinct problems using the same underlying algorithm. An open-source implementation of the state-of-the-art algorithm Proximal Policy Optimization is adopted to carry out the training process of a deep neural network for trajectory design of interplanetary missions [7]. In that same work, the stochastic optimal control problem is recast as a time-discrete Markov decision process, to comply with the standard formulation required by reinforcement learning [7]. This recasting step, turning a continuous-time optimal control problem into a sequence of discrete decisions with states, actions and rewards, is a general prerequisite for applying RL algorithms to any guidance problem, not only the interplanetary case where it is described.

Image-based approaches extend RL guidance to cases where the policy must act on visual input rather than a clean state estimate. Image-based deep reinforcement meta-learning has produced a closed-loop guidance policy that reaches errors in the order of meters across different scenarios, including when the environment is only partially observed, for autonomous lunar landing [3]. A related image-based meta-reinforcement learning guidance system, applied to the different problem of guiding an asteroid impactor, correctly drove the spacecraft toward the final impact point in more than 98% of 500 test scenarios [8]. Both results describe closed-loop guidance behaviour under uncertainty about the true state, which is directly relevant to a small-scale rocket that must also rely on imperfect state estimation, though neither claim is about a rocket landing on a launch pad or barge.

Standard RL algorithms are not automatically reliable for this class of goal-oriented landing task, and this is an important caution for anyone assuming that any RL algorithm will simply work out of the box. PPO, SAC and DSAC are reported to fail in goal-oriented rocket landing scenarios because they rely on extensive exploration, and the likelihood of randomly reaching the goal diminishes exponentially over time [6]. This failure mode is why specialised training schemes, described in the next section, have been proposed specifically for rocket landing control, rather than relying on off-the-shelf PPO, SAC, or DSAC training alone.

Random Annealing Jump Start (RAJS) is a training method designed to overcome the exploration failure of standard RL algorithms in goal-oriented rocket landing. In RAJS, a guide policy navigates the environment for a guide horizon, and then an exploration policy takes charge to complete the remaining steps [6]. The guide horizon is sampled from a uniform distribution whose upper bound anneals to zero based on performance metrics [6]. This design lets the exploration policy start closer and closer to the goal as training proceeds, since early in training the guide policy covers most of the trajectory, and as performance improves the guide policy's share of each episode shrinks toward nothing, handing progressively more of the task to the exploration policy under training.

The mechanism directly targets the exponential exploration problem described for standard PPO, SAC and DSAC. Instead of asking the exploration policy to stumble onto a rare goal state from a fully random start, RAJS gives it a running start from a state reached by a competent guide policy, and only gradually withdraws that head start. This design choice is a direct response to the stated cause of failure in standard RL algorithms, namely that the likelihood of randomly reaching the goal diminishes exponentially over time [6].

On a high-fidelity rocket model, RAJS raised the success rate of rocket landing control from 8% with a baseline controller to 97% [6]. This is a large improvement, but it is reported for the specific high-fidelity rocket model used in that study, not for a small-scale 10 kg, 100 m apogee vehicle. A builder targeting a small-scale rocket should treat this number as evidence that the training scheme works on the vehicle class it was tested on, and should not assume the same 8% to 97% jump would occur on a different vehicle without re-running the experiment.

The RAJS method was also validated through extensive evaluation and Hardware-in-the-Loop testing, which affirmed the effectiveness, real-time feasibility, and smoothness of the proposed controller [6]. Hardware-in-the-Loop testing is a meaningful step toward real deployment because it exercises the controller against real timing and interface constraints, not just a simulated clock. However, this validation was carried out on the high-fidelity rocket model used in that paper; the claim does not state that this Hardware-in-the-Loop result was obtained on, or is transferable to, a small-scale 10 kg vehicle. A builder considering RAJS as a fallback controller for a small-scale rocket should note this mismatch in vehicle scale and fidelity explicitly rather than assume the real-time and smoothness result carries over.

Taken together, RAJS is best read as a solution to a specific known failure of standard RL training in this domain, evaluated on one vehicle model, rather than as a general guarantee of performance across rocket scales. It is a training-time fix, not a guidance-law change, and it can in principle be paired with any of the reward-shaping or discount-rate tuning ideas reported elsewhere for RL guidance.

Convex optimization is the established classical approach to powered descent guidance. Lossless convexification of nonconvex cost, dynamics, and control constraints is often applied to the 3-DoF powered descent guidance problem [4]. Convex optimization has been considered a prime candidate for on-board autonomous guidance applications because of its deterministic behaviour, its global optimality guarantees, its ability to provide certificates of convergence and infeasibility, and the availability of efficient interior point method (IPM) algorithms [14].

These properties, determinism, optimality guarantees, and convergence certificates, are exactly what an RL policy does not offer by construction, since a trained neural network provides no formal guarantee that its output is optimal or even feasible for a new, unseen scenario. This is a structural difference between the two families rather than a measured comparison, and it is worth stating plainly for a builder deciding which family to trust for a safety-critical landing phase.

The 6-DoF version of the powered descent problem is harder than the 3-DoF version. It is known to be challenging to solve quickly and reliably because the problem is nonlinear and non-convex, the discretization scheme heavily influences solution validity, and reference trajectory initialization determines whether the algorithm converges or diverges [4]. An earlier paper describing successive convexification for 6-DoF Mars rocket powered landing with free-final-time states the same three difficulties in near-identical terms: solving the problem quickly and reliably is challenging because it is nonlinear and non-convex, because the validity of the solution is heavily dependent on the accuracy of the discretization scheme, and because it can be difficult to select a suitable reference trajectory to initialize an iterative solution process [14].

That successive convexification work also reports that, through the use of Interior Point Method (IPM) solvers, this sequence of convex subproblems can be solved quickly and reliably, thus enabling higher-fidelity real-time guidance for rocket powered landings on Mars [14]. This is a positive result specific to the Mars powered-landing setting and to IPM-based solvers, and it should be read as evidence that the successive-convexification approach, when paired with IPM solvers, meets a real-time requirement in that setting, not as a general speed guarantee independent of solver choice.

Separately, a convex programming method for rocket powered landing with an angle-of-attack constraint showed, in numerical simulations, that it can find a feasible landing trajectory where the angle-of-attack constraint is satisfied, with better convergence performance compared with a traditional linearization method [10]. This is a comparison against a linearization baseline within the convex-optimization family, not against an RL policy, and it demonstrates that convex methods themselves continue to be refined against earlier convex or linearized techniques.

Precision landing under real disturbances adds further difficulty beyond the algorithmic challenges above. Endoatmospheric landing is subject to disturbing conditions including engine thrust fluctuation, aerodynamic coefficient uncertainty, atmospheric density perturbation, and wind disturbance [5]. This is the backdrop against which any guidance law, learned or classical, must be judged if it is meant for a real, flying small-scale rocket rather than only a clean simulation, since a small, light vehicle is likely to be more sensitive to wind and thrust variation than a large launch vehicle, even though no claim states this comparison directly.

A separate line of work does not pit RL against convex optimization but instead uses learning to make convex optimization faster or more reliable. T-PDG improves solution times by up to an order of magnitude compared to lossless convexification (LCvx) for the 3-DoF minimum-fuel powered descent guidance problem [4]. This is a timing comparison between a learning-augmented method and a classical baseline, on a specific 3-DoF problem formulation, not an accuracy-versus-accuracy comparison of an RL policy versus LCvx.

The mechanism behind this speed-up is prediction of which constraints will be active at the optimal solution. T-SCvx learns to predict the set of tight or active constraints at the optimal control problem's solution, creating a minimal reduced-size problem initialized with only the tight constraints, then uses the solution to warm-start the direct optimization solver [4]. Warm-starting means the solver starts closer to the answer, which is why solve times drop; the reduced problem size, containing only the predicted tight constraints instead of the full constraint set, is what allows the solver to converge faster than it would on the full lossless-convexification formulation.

A related method, TOAST, follows a similar warm-start philosophy but is evaluated on different problems. Through numerical experiments on a Lunar rover problem and a 3-degrees-of-freedom Mars powered descent guidance problem, TOAST outperforms benchmark approaches in terms of both computation times and network prediction constraint satisfaction [11]. Again, this is a comparison against benchmark warm-start or optimization approaches, not against a pure end-to-end RL guidance policy, and the two test problems, a Lunar rover and a 3-DoF Mars powered descent case, are distinct from both the 6-DOF planetary lander used for the RL guidance policy in [1] and from any small-scale rocket.

For a builder, the practical lesson from this family of methods is that learning and convex optimization are not mutually exclusive. A system can use a learned network purely to accelerate or stabilise a convex solver's convergence, keeping the solver's guarantees, rather than replacing the solver's decision-making with a network end to end as in the RL guidance policies described earlier. This hybrid path is worth considering as a third option alongside pure RL and pure convex optimization when designing the comparison experiment this note is ultimately about.

Precision and robustness to dispersion are central requirements for any small-scale reusable rocket, since real hardware never behaves exactly like its simulation model. One proposed approach is a novel online trajectory dispersion control method based on a Parameterized Optimal Feedback Guidance Law (POFGL) [5]. This is a guidance law aimed specifically at controlling how much a rocket's actual trajectory disperses from its planned one under disturbance, rather than a general-purpose optimizer or learned policy, and it is presented as an online method, meaning it is meant to run continuously during flight rather than being solved once before launch.

A known limitation of feedback guidance laws of this kind is behaviour near the end of the descent. As the time-to-go tends to zero, the sensitivity of the generated guidance command trajectories will significantly increase and inevitably tend to infinity [5]. This is a structural property of the guidance law near touchdown, and any implementation must handle it explicitly, for example by switching guidance mode or capping commanded sensitivity as time-to-go approaches zero, though the claim itself does not prescribe which specific mitigation to use.

This dispersion problem is stated against the backdrop of the disturbances already noted for endoatmospheric landing: engine thrust fluctuation, aerodynamic coefficient uncertainty, atmospheric density perturbation, and wind disturbance [5]. These are exactly the kinds of disturbances a small-scale rocket, which is lighter and more sensitive to wind and thrust variation than a large launch vehicle, would be expected to encounter in an outdoor test, even though no claim makes this scale comparison directly.

None of the claims in this set report a landing-accuracy or fuel-consumption number for POFGL that could be placed alongside an RL policy's number under the same test conditions. The value of this section for a builder is the reminder that guidance laws, whichever family they come from, must be checked specifically for behaviour as time-to-go approaches zero and under realistic disturbance models, not only in a nominal, disturbance-free run, since a guidance law that performs well in a clean simulation may still fail near touchdown or under wind and thrust disturbances that were not modelled.

To keep this note honest, here is every quantitative or comparative result in the claim set, stated with its own baseline and conditions, without merging results from different papers.

• RAJS raised rocket landing control success rate from 8% with a baseline controller to 97%, on a high-fidelity rocket model, using RL [6].
• T-PDG improves solution times by up to an order of magnitude compared to LCvx, for the 3-DoF minimum-fuel powered descent guidance problem [4].
• TOAST outperforms benchmark approaches in computation time and network prediction constraint satisfaction, on a Lunar rover problem and a 3-DoF Mars powered descent guidance problem [11].
• The convex programming method with an angle-of-attack constraint has better convergence performance than a traditional linearization method, in numerical simulations, and finds a feasible landing trajectory satisfying that constraint [10].
• An image-based meta-reinforcement learning guidance system correctly drove a spacecraft to the final impact point in more than 98% of 500 test scenarios, for an asteroid impactor problem [8].
• An image-based deep reinforcement meta-learning policy reached errors in the order of meters across different scenarios, including partial observability, for lunar landing [3].
• The RAQResNet model achieved a validation loss approximately 300 times lower than a standard architecture with an equal number of trainable parameters, and 50 times lower than a standard architecture with twice the number of trainable parameters [2].

None of these results compare an RL guidance policy and a classic convex guidance law on the same vehicle and the same accuracy or fuel metric. Each number belongs to its own paper's benchmark, sample size, and setting, and mixing them into a single ranking would misrepresent the evidence. A builder should treat each of these numbers strictly as a report about the specific method, vehicle, and test scenario named alongside it, and should not average or combine them across papers to produce a claimed ranking between RL guidance and convex-optimization guidance in general.

Two pieces of infrastructure work, while not guidance methods themselves, are relevant background for a team building and testing RL guidance policies. The Open Ant is a physical variant of the commonly used Gymnasium Ant environment, along with a simulation [15]. The Open Ant robot body is designed to look like the widely used Gymnasium Ant environment [15], and it is designed to be built and maintained by AI researchers without backgrounds in robotics engineering [15]. This second point matters for a small rocket-guidance team: a platform explicitly designed for non-robotics researchers lowers the engineering barrier to running physical RL experiments, which is directly relevant if a team without deep aerospace hardware experience wants to test a guidance policy on real hardware rather than only in simulation.

Mechanically, the Open Ant uses an overhead webcam to track fiducial markers to compute reward signals and the heading vector of the ant [15]. The robot is connected by cables to AC power and to an external computer where the agent is running [15]. This means the reward computation and the agent's decision-making both happen off-board, relying on external sensing and compute rather than fully autonomous onboard systems, which simplifies the hardware a research team needs to build compared to a fully self-contained robot.

On the learning side, competent walking policies can be learned from scratch in approximately one hour directly from the physical robot's experience, for SARSA(λ) and Soft Actor-Critic (SAC) [15]. Separately, the Open Ant comes with a MuJoCo simulation, which can be used to learn competent policies [15], and policies learned in simulation transfer to reality for the Open Ant platform [15]. Together these two results describe two independent routes to a working policy on this platform: training directly on the physical robot in about an hour, or training in the MuJoCo simulation and transferring the result to the real robot. Neither of these results is about rockets or guidance, but the pattern, a matched simulation-and-hardware pair with demonstrated sim-to-real transfer, is exactly what a team building a small-scale rocket testbed to answer the RL-versus-convex-optimization question would need.

Separately, Gym-μRTS is described as a fast-to-run RL environment for full-game real-time strategy research [17]. This is unrelated to rocket guidance in domain, but it is evidence that fast, purpose-built RL environments are valued in the RL research community generally as a way to make iteration cheap, which is a consideration for any team designing a rocket-guidance training environment from scratch. Neither the Open Ant nor Gym-μRTS answers anything about landing accuracy or fuel efficiency; they are included here because they show what matched simulation-hardware infrastructure and fast training loops look like in other domains, which is exactly the kind of infrastructure a rocket-guidance comparison project would need to build for itself.

A few further claims describe RL applied to problems adjacent to rocket landing guidance, useful for a builder deciding which RL techniques to borrow. In liquid rocket engine transient control, a deep reinforcement learning controller achieved the highest performance compared to carefully tuned open-loop sequences and PID controllers [13]. The same study showed that the learned policy can reach different steady-state operating points and adapt convincingly to changing system parameters [13]. This is a different control problem, engine transient control rather than descent guidance, but it is evidence from the same broad RL-for-rocket-systems literature that a learned controller can beat classical open-loop and PID baselines in at least one rocket subsystem, and it shows an RL controller adapting online to changing system parameters, a property distinct from, but related in spirit to, the noise and parameter-uncertainty robustness reported for the 6-DOF descent guidance policy.

On the RL algorithm-design side, existing approaches to curriculum or task selection in RL typically approximate task differences using agent-centric signals such as regret or value disagreement, but these signals are inherently policy-dependent and sensitive to noise [12]. Causal-Paced Deep Reinforcement Learning (CP-DRL) is proposed as a framework that addresses this limitation by approximating the structural causal model (SCM) difference between tasks using only interaction data [12]. This is general RL methodology, not specific to rocket guidance, but it is relevant if a builder plans to train a guidance policy across a curriculum of increasingly difficult landing scenarios, for instance progressively larger initial dispersions or wind disturbances, and wants a principled, noise-robust way to choose which scenario to train on next rather than relying on regret or value-disagreement signals that the claim states are sensitive to noise.

Finally, on the real-world recovery problem this note is ultimately about, one design paper on retro-propulsive vertical landing of a reusable rocket stage states plainly that one of the challenges lies in the recovery guidance, navigation and control (GNC) strategy and algorithms, in particular those of the powered-landing phase, which must enable a precise landing with low fuel margins and significant dispersions [9]. This is a framing of the problem's difficulty, not a solution or a number, and it applies to full-scale reusable rocket recovery. It is stated here as background for why the guidance question matters at all, independent of scale: the tension between precision, low fuel margins, and significant dispersions is the same tension that any small-scale RL-versus-convex comparison experiment must be designed to probe.

None of these three claims bear directly on the small-scale RL-versus-convex comparison, but each narrows down what a builder should expect to borrow, and from where: control techniques and online-adaptation evidence from engine transient work, curriculum methods from CP-DRL, and problem framing from recovery GNC design.

The central limitation is direct: this evidence set contains no experiment in which an RL guidance policy and a classic convex-optimization guidance law are run on the same small-scale rocket, under the same test scenarios, and scored on the same accuracy and fuel metrics. Every quantitative result above belongs to its own paper's vehicle model, benchmark, and baseline, and the numbers cannot be combined into a ranking between the two families of method without going beyond what any single claim states.

A second limitation is scale mismatch. The RL results reported here come from a 6-DOF planetary lander simulation [1], a high-fidelity rocket model [6], asteroid impactor and lunar landing scenarios [8][3], none of which is stated to be a 10 kg, 100 m apogee small-scale rocket. Likewise the convex-optimization results come from Mars powered landing and 3-DoF or 6-DoF formulations [14][4][10], not from a small-scale vehicle. Applying either family's reported numbers to a small-scale rocket is an extrapolation the claims do not support, and a builder must re-run the relevant experiments on the actual vehicle of interest before drawing conclusions.

A third limitation concerns training reliability. Standard RL algorithms, including PPO, SAC, and DSAC, are reported to fail in goal-oriented rocket landing scenarios due to reliance on extensive exploration [6]. This means a builder cannot simply apply an off-the-shelf PPO implementation, as used elsewhere for interplanetary trajectory design [7], and expect it to succeed at landing without a specialised training scheme such as RAJS, or reward shaping choices such as the discount-rate tuning reported for the 6-DOF planetary lander [1].

A fourth limitation concerns robustness claims being scenario-specific. Hardware-in-the-Loop testing for RAJS affirmed real-time feasibility and smoothness on the high-fidelity rocket model tested [6], and this should not be read as evidence for a differently scaled or differently fidelity vehicle. Likewise, the structural sensitivity problem as time-to-go approaches zero, reported for POFGL [5], and the list of endoatmospheric disturbances [5], both apply to the specific guidance law and flight regime discussed there, and a builder must re-verify these behaviours for any new guidance law and vehicle rather than assume they transfer.

A fifth, more general limitation is that convex-optimization guarantees and RL performance reports are not measured on comparable terms even in principle within this evidence set. Convex optimization's advantages are described as deterministic behaviour, global optimality guarantees, and certificates of convergence or infeasibility [14], which are properties of the algorithm's structure, not numbers measured on a test set. RL's reported advantages, accuracy, fuel efficiency, and robustness to noise [1], or success rate on a specific vehicle [6], are empirical numbers from specific test scenarios. A fair comparison would need to measure both families on the same empirical footing, for example landing error and fuel consumed over an identical scenario set, which no claim here provides.

The following procedure is for a team that wants to actually produce the missing head-to-head comparison for a small-scale rocket, using the methods described in the claims above as building blocks. Every design choice below beyond what the claims state (specific parameter values, integrator types) is illustrative only, chosen to make the procedure concrete, and is not asserted as coming from any cited source.

1. Define the vehicle and scenario set. Fix the vehicle: 10 kg dry mass, 100 m apogee, with mass, thrust, and aerodynamic properties specified. Define a fixed set of initial conditions (position, velocity, attitude dispersions) that both guidance methods will be tested against, since none of the cited papers share a common benchmark you can reuse directly.
2. Build the simulation. Implement rigid-body dynamics for the vehicle, in 3-DoF or 6-DoF depending on which formulation you choose for the classical baseline. No cited claim specifies an integrator or simulation software; choose one appropriate to your accuracy and speed needs and document it explicitly for reproducibility.
3. Implement the classical baseline first. Implement lossless convexification (LCvx), often applied to the 3-DoF powered descent guidance problem [4]. If you need 6-DoF, be aware this is reported as harder to solve quickly and reliably, because of the nonlinear and non-convex nature of the problem, the discretization scheme's influence on solution validity, and the reference trajectory's role in determining convergence or divergence [4][14].
4. Verify the baseline against known difficulties. Test the convex solver's sensitivity to discretization scheme and to reference-trajectory initialization, since both are named as causes of convergence or divergence [4][14]. If using an Interior Point Method solver, check whether it meets a real-time solve-time budget, as reported for the successive-convexification approach on Mars powered landing [14]. Record solve time and success rate over your scenario set before touching any learned method.
5. Choose an RL algorithm and formulate the MDP. Recast the guidance problem as a time-discrete Markov decision process, following the same recasting done for a related stochastic optimal control problem [7]. Use Proximal Policy Optimization (PPO) as a first algorithm, following its use for state-to-thrust mapping in 6-DOF planetary powered descent and landing [1]; be prepared for it to fail on its own in a strict goal-oriented landing task, since standard RL algorithms including PPO are reported to fail here due to exploration difficulty [6].
6. If PPO alone fails, add a guided-exploration training scheme. Implement a training scheme in the spirit of Random Annealing Jump Start: a guide policy that navigates the environment for a sampled guide horizon, handing off to an exploration policy for the rest of the episode, with the horizon's upper bound annealing to zero as performance improves [6]. This is the method reported to raise success rate from 8% to 97% on a high-fidelity rocket model [6]; treat that number as evidence the technique works on that vehicle, not as a guarantee for your own.
7. Tune the reward structure. Consider using different discount rates for terminal reward versus shaping reward, since this is reported to enhance optimization performance for a 6-DOF planetary lander [1]. Validate this choice independently for your own reward formulation, since the claim reports it for a specific study's reward design.
8. Add noise and parameter-uncertainty robustness testing. Run both the trained RL policy and the convex baseline against injected sensor noise and perturbed vehicle parameters, since robustness to noise and parameter uncertainty is a property reported for the 6-DOF RL guidance system [1]. Report success rate and landing error under each perturbation level for both methods, on the same perturbation set, so the comparison is apples-to-apples.
9. Measure landing accuracy and fuel use for both methods, side by side. For each of the fixed scenarios from step 1, record final position and velocity error and total propellant consumed, for both the RL policy and the convex baseline. This is the core experiment the surveyed literature does not provide, so design it carefully: same scenarios, same disturbance models, same metrics, same vehicle.
10. Consider a hybrid warm-start approach as a third condition. Optionally implement a learned warm-start for the convex solver, in the spirit of T-PDG or T-SCvx, which learn to predict the set of tight or active constraints at the optimal solution and use this prediction to build a reduced-size problem that warm-starts the solver [4]. Compare its solve time against plain LCvx, since T-PDG is reported to improve solution times by up to an order of magnitude compared to LCvx for the 3-DoF minimum-fuel problem [4], and include this hybrid as a third condition alongside pure RL and pure convex optimization.
11. Stress-test near touchdown. If using any feedback-style guidance law, check behaviour as time-to-go approaches zero, since the sensitivity of commanded trajectories is reported to increase sharply and tend to infinity in that regime for one such law [5]. Cap or switch guidance mode near touchdown as needed, and document the switch explicitly since it is not itself validated in the cited claim.
12. Report results honestly. State the vehicle, scenario set, and metrics used, and do not extrapolate your small-scale results to full-scale rockets or vice versa, since none of the cited results were obtained on a 10 kg, 100 m apogee vehicle. If your RL policy needs Hardware-in-the-Loop testing before flight, note that such testing has been used elsewhere to check real-time feasibility and smoothness for a different, higher-fidelity rocket model [6], and treat any transfer of that result to your smaller vehicle as unverified until you run it yourself.

function pipeline():
    baseline_result = run_LCvx(vehicle, scenarios)            # per [4]
    rl_policy = train_PPO(vehicle_MDP)                          # per [1],[7]
    if rl_policy.success_rate is low:
        rl_policy = train_with_guided_exploration(vehicle_MDP)  # per [6]
    rl_result = evaluate(rl_policy, scenarios)
    hybrid_result = run_warm_started_solver(vehicle, scenarios) # per [4], optional
    compare(baseline_result, rl_result, hybrid_result)
    # this three-way comparison is the missing evidence, not found in any cited claim