{"id":"82a07588-8e6f-495f-aa67-f3d69a1004ed","arxiv_id":"2501.00915","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A diffusion model trained on 6DoF powered-descent solutions can be composed at inference time with glideslope and risk-map energy functions to generate constrained, multi-modal landing trajectories without retraining.","lead":"The authors use a type of AI generative model called a diffusion model to create spacecraft landing trajectories, and they show that combining separate models at the moment of use lets the same trained model handle new landing-site constraints without retraining. The approach is aimed at making powered-descent guidance faster and more flexible for future onboard spacecraft autonomy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reverse-diffusion-only composition is an uncontrolled approximation: the paper's Eq. (13) bounds score error, not constraint violation, and the reported gains rely on hand-tuned energy weights with no sensitivity or generalization analysis.","rationale":"The reader's weakest assumption identifies the same technical soft spot: composition via energy addition/negation, carried out with reverse diffusion only, is not guaranteed to sample the desired constrained distribution. This is load-bearing because every claimed contribution—inference-time glideslope satisfaction, multi-landing-site selection, and the 'no retraining' benefit—depends on that approximate sampler. The paper's Eq. (13) is a bound on score error, not on the probability of constraint violation, and the energy forms in Eqs. (31) and (34) are explicitly hand-crafted with tuned weights. The reported violation rates are point estimates from a single configuration, so they cannot establish that the approximation holds across the problem variations the paper claims to address. A concrete MCMC comparison and weight-sensitivity check would settle whether the reported constraint satisfaction is due to the composition mechanism or to the specific tuned constants. This concern reinforces, rather than changes, the reader's CONDITIONAL verdict; the paper's statistical comparisons of the base trajectory distribution are useful evidence, but the chief compositional claim remains unverified. I therefore recommend leaving the verdict unchanged while requiring the additional experiments before the claim can be accepted.","tokens_in":18599,"tokens_out":3609,"duration_ms":32710,"concrete_test":"Run the glideslope and landing-site composition experiments on the released (or provided) base model, drawing N=1000 samples in each of three conditions: (i) reverse diffusion only, as in the paper; (ii) reverse diffusion followed by 50 ULA MCMC steps per diffusion timestep on the same composed energy; (iii) reverse diffusion with alpha2 (and lambda for the risk map) varied by ±50%. If condition (ii) materially reduces violation rates below the reported 4.5% and <7%, or condition (iii) changes violation rates by more than the reported baseline differences, then the paper's headline results are artifacts of the approximate sampler and its tuned weights.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that a single trained dynamics diffusion model can be composed at inference time to generate constraint-abiding, multi-modal trajectories for new problems—rests on the assumption that negation composition with reverse diffusion alone samples close enough to the composed distribution. Section III.B explicitly states that reverse diffusion 'does not sample from the exact composed distribution,' and the cited Eq. (13) is an upper bound on score error, not a bound on constraint-violation probability. The glideslope energy (Eq. 31) and the obstacle/risk energy (Eq. 34) contain hand-set constants (alpha1=1.3, alpha2=0.3; lambda, tau, w_k, and s(t) in Eq. 36) that are tuned for the demonstration. Reported improvements (4.5% vs 10% glideslope violations; <7% vs 11% high-risk landings) are single-configuration point estimates with no error bars, no sensitivity analysis, and no variation of the initial-condition ranges in Table 2 or of the obstacle layout. Since the entire 'no retraining' benefit is realized through this approximate sampler, the method's generality is unestablished: if approximation error grows outside the tuned setting, the composed model no longer satisfies constraints and the claimed advantage disappears.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper trains a denoising diffusion probabilistic model on 6-degree-of-freedom powered-descent guidance trajectories generated by a successive convexification solver, with random initial conditions and random final horizontal positions. It then uses energy-based composition, specifically negation composition with hand-crafted glideslope and risk-map energy functions, to modify the generated distribution at inference time, and uses inpainting to enforce equality constraints such as waypoints. The reported results are distributional comparisons between diffusion-generated and optimizer-generated trajectories, a glideslope-violation comparison (4.5% vs. 10%), a qualitative inpainting demonstration with multimodality, and a risk-map comparison (<7% vs. 11% high-risk landings). The paper claims this is the first generative-composition approach to high-dimensional multi-landing-site trajectory generation and that the generated samples are useful as initial guesses for powered-descent guidance.","tokens_in":18917,"tokens_out":7774,"duration_ms":71795,"significance":"If the claims hold, this would be a practically valuable contribution: a single dynamics model could be reused for multiple mission constraints and landing-site selections without retraining, and the demonstrated multimodality is genuinely useful for sampling-based warm starting. The paper is honest about its main limitation, namely that reverse diffusion does not sample the exact composed distribution, and the inpainting portion is exact by construction. The risk-map composition is a clear and nontrivial demonstration. However, the experimental support is currently narrow: all constraint-satisfaction claims come from single configurations with hand-tuned energy weights, no error bars, no sensitivity analysis, no warm-start solver study, and no evaluation outside the training distribution. The significance is therefore conditional on substantially strengthening the empirical evaluation.","major_comments":[{"comment":"The boundary conditions fix the full final position vector to zero, but the accompanying text states that 'only the z component of the final state is constrained to be zero, while the x and y components are sampled,' and Table 2 lists r_x,N and r_y,N as uniform on [-4,4]. This internal contradiction affects the definition of the multi-landing-site dataset and the diffusion model's training target; please clarify which boundary conditions were actually used to generate the training data.","section":"Section IV.A, Eqs. (18e), (23e), and Table 2"},{"comment":"The abstract claims that the method efficiently adapts out-of-distribution data and problem variations in a few-shot framework and that the samples enable computationally efficient trajectory generation, but the paper contains no out-of-distribution test (e.g., initial conditions outside the ranges in Table 2, different glideslope angles, or unseen obstacle layouts) and no warm-start study in which diffusion samples initialize SCvx or an NLP solver. Section V.A reports only raw sampling times (1.17 min for 1000 diffusion samples versus about 5.67 min for 100 optimizer solutions). Because the warm-start benefit is a central claimed benefit, this experiment or a tempering of the claim is necessary.","section":"Abstract and Section V"},{"comment":"The reported glideslope improvement (4.5% vs. 10%) and high-risk-landing improvement (<7% vs. 11%) are single point estimates from one configuration of hand-set weights (alpha_1=1.3, alpha_2=0.3, lambda_penalty, lambda, tau, w_k, s(t)), with no confidence intervals, no seed variation, and no sensitivity analysis. Because Section III.B explicitly states that reverse diffusion does not sample the exact composed distribution, and Eq. (13) bounds score error rather than constraint-violation probability, the manuscript needs violation statistics over repeated sampling runs and over ranges of these parameters and problem parameters before the 'constraint-abiding' claim is supported.","section":"Section V.B, Eq. (31); Section V.D, Eq. (34); Section III.B"},{"comment":"The derivation of the score loss, the score-error bound in Eq. (13), the cosine beta scheduling, and the initialization of the trained model all come from reference [14], which is authored by the first author of this paper. The manuscript does not state whether [14] is publicly or archivally available, and it does not include the derivation or the initialization hyperparameters. Since these are load-bearing for the trained model and the theoretical justification of composition, please either provide the required derivation and training details in an appendix or confirm that [14] is accessible and version-stable.","section":"Section III.A.1 and Section V.A.1"}],"minor_comments":[{"comment":"The entry defining alpha_t as the product of beta_t is inconsistent with the sampling equation (7), where alpha_t appears as the product of (1 - beta_t); please correct the definition.","section":"Nomenclature"},{"comment":"The caption says 'multi-model structure' but should say 'multi-modal structure.'","section":"Figure 2 caption"},{"comment":"The claim of 'reducing constraint violations by 40%' does not follow exactly from 11% to less than 7% (approximately 36%); please report the measured values and the calculation.","section":"Section VI"},{"comment":"The score-error bound is a single inequality split across two numbered equations; renumbering as one equation would avoid the impression that there are two separate bounds.","section":"Eqs. (13)-(14)"},{"comment":"The phrase 'transported convolutional layers' should be 'transposed convolutional layers.'","section":"Section V.A"},{"comment":"The claim that inpainting generates trajectories that 'exactly meet all equality constraints' would be stronger with a numerical error table (e.g., max absolute constraint error over a sample set), since Figure 17 alone is qualitative.","section":"Section V.C"},{"comment":"The U-Net description lists only the conv2_d weight shapes; providing the full architecture details, including skip connections, normalization, and any attention layers, or a pointer to code, would improve reproducibility.","section":"Section V.A, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the direction is appropriate, but the novelty boundary relative to reference [14] needs careful checking: the first author of this manuscript is also the first author of [14], and the present paper appears to build directly on a model initialized from that work. The unresolved contradiction in the final-position boundary conditions (Eqs. (18e)/(23e) versus Table 2) should be resolved before further review. The main requested experiments (warm-start study, sensitivity analysis, out-of-distribution evaluation) are within the scope of a revision and would substantially increase confidence in the central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Nothing wrong with the core trick. The paper takes the compositional diffusion machinery from Du and Kaelbling and the authors' own TrajDiffuser [14], and shows it carries over to 6-DoF powered descent with multi-landing-site selection. The distillation into a 2D array of states and controls and the EBM-based negation composition with a risk map is a clean way to get multi-modal trajectories to different landing sites without retraining. The distributional comparison against SCvx, with 1,000 samples each and means and standard deviations on positions, velocities, quaternions, angular rates, and thrusts, is the kind of honest baseline check I like to see. The inpainting results are exact by construction, and the waypoint-conditioning figure makes the point visually.\n\nWhat is actually new: to my knowledge, applying density composition to the multi-landing-site selection problem, with the risk map entering as a negation energy with per-point weights and a sinusoidal schedule, is new, as is the waypoint inpainting demonstration. I also credit the authors for not overselling the algebra; Section III.B explicitly says reverse diffusion does not sample the exact composed distribution and defers MCMC to future work.\n\nWhere I agree with the stress-test note: the abstract and conclusion reach beyond the evidence. 'Few-shot' and 'out-of-distribution' appear, but there is no OOD experiment. The test scenarios are the same distribution as training, just with new constraints and risk maps. There is no warm-start comparison against an optimizer, which is the stated use case. The headline numbers, 4.5% vs 10% glideslope violations and less than 7% vs 11% high-risk landings, are single point estimates with no error bars and no sensitivity study over alpha_1, alpha_2, lambda, tau, w_k, or s(t). Equation (13) bounds score error, not constraint-violation probability. These are real limitations, but not fatal ones for a first demonstration. A single-configuration result is enough to motivate the approach; it just does not support the 'generalizable' language.\n\nThe circularity concern is fair but should not be overstated. The paper leans heavily on [13,14] for the ULA interpretation, the score error bound, the architecture, and even the model initialization. None of that is independently verified here. That is a citation-pattern issue, not misconduct; [14] is genuinely prior work on this exact trajectory class.\n\nWho this is for: people building sampling-based warm-starters for onboard guidance, and researchers applying compositional diffusion to constrained trajectory generation. It deserves a serious referee, not a desk reject. I would send it back for major revision: soften the abstract to what is actually shown, or add the missing OOD, warm-start, and sensitivity experiments and release code and data.","headline":"A clean engineering demonstration of compositional diffusion for 6-DoF landing-site selection, but the 'few-shot/OOD' claims outrun the experiments and the inference-time composition is only validated on hand-tuned single-configuration demos.","tokens_in":837,"tokens_out":1981,"would_cite":true,"duration_ms":44447,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that composing a trajectory diffusion model with risk maps and constraint energies at inference time generates constraint-abiding, multi-modal spacecraft landing trajectories without retraining.","keywords":["diffusion models","trajectory generation","powered descent guidance","compositional generative modeling","energy-based models","multi-landing site selection","successive convexification","spacecraft autonomy"],"falsifier":"Re-run the glideslope composition experiment and measure the fraction of trajectory segments violating the constraint for the base and composed models; if the composed model does not reproducibly cut the violation rate below the base model's roughly 10%, or if the reduction disappears when the penalty weight $\\lambda_{\\text{penalty}}$ is perturbed, the composition assumption is not doing the claimed work.","tokens_in":18407,"feed_emoji":"🚀","tokens_out":8524,"duration_ms":70681,"temperature":0.7,"pith_summary":"Current machine-learning methods for spacecraft trajectory generation learn a single parameter-to-solution mapping, so any change in mission constraints forces a new dataset and a new model. This paper argues that diffusion models avoid that limitation by representing the trajectory distribution as a probability density that can be composed, at inference time, with other densities encoding costs and constraints. The central demonstration is a 6-degree-of-freedom powered descent guidance model trained on successive-convexification solutions: negation composition with a differentiable glideslope energy cuts constraint violations, inpainting enforces waypoint and boundary equality constraints exactly, and composing with a low-risk landing-site map steers samples away from high-risk sites. If correct, a single learned dynamics model can generate the multi-modal, constraint-abiding initial guesses that warm-start trajectory optimization, removing the need to retrain for each new landing site or constraint.","feed_headline":"One diffusion model adapts to new landing sites without retraining","feed_subtitle":"Composing the model with risk maps and constraint energies produces safe, multi-modal descent guesses at sampling time.","key_machinery":"The central object is the compositional diffusion model, which treats a trained trajectory diffusion model as an energy-based model: the learned energy $E_\\theta(\\mathbf{x},t)$ is proportional to the negative log-likelihood, and the score function is $\\boldsymbol{\\epsilon}_\\theta(\\mathbf{x}_t,t)=\\nabla E_\\theta(\\mathbf{x}_t,t)/\\sqrt{1-\\bar\\alpha_t}$. Composition combines independent models through product $p_\\theta(\\mathbf{x})\\prod_k g_k$, negation $p_\\theta(\\mathbf{x})/h(\\mathbf{x})^\\alpha$, or inpainting with a Dirac-delta constraint $h(\\tau)=\\delta_{g_t(s_0,a_0,\\dots)}$. Constraint energies are hand-crafted differentiable functions, e.g., the glideslope energy $E_{\\text{constraint}}=-\\lambda_{\\text{penalty}}\\,\\text{violation}^2$ and the obstacle energy built from a normalized risk map with bilinear interpolation and a sinusoidal time scaling $s(t)=\\sin(\\pi t/T)$. These energies are summed with weights ($\\alpha_1=1.3$, $\\alpha_2=0.3$) to form the composed energy, whose gradient drives reverse diffusion; the reverse process is annealed ULA MCMC without full MCMC iterations, so samples approximate rather than exactly match the composed distribution.","core_discovery":"The paper's discovery is that a diffusion model trained only on unconstrained 6-DoF powered-descent trajectories, produced by successive convexification (SCvx), already encodes the dynamics well enough that new mission requirements can be imposed at sampling time by composing probability densities. Formally, the composed distribution is $p_\\theta(\\mathbf{x}) \\propto \\prod_i p^i_\\theta(\\mathbf{x}\\mid\\text{env})\\prod_j c_j\\prod_k g_k$, where $p^i_\\theta$ are independently trained trajectory diffusion models, $c_j$ are cost distributions, and $g_k$ are constraint distributions. Using energy-based formulations, the glideslope constraint is encoded as a differentiable violation energy under negation composition, reducing glideslope violations to below 5% of trajectory segments versus roughly 10% for the base model; inpainting forces waypoint equality constraints exactly while preserving multi-modality; and composition with a risk map reduces landings in high-risk regions by about 40% (from 11% to under 7%). The authors claim this is the first use of generative composition and multi-modality for high-dimensional multi-landing-site trajectory generation, with all samples produced without retraining and used as initial guesses for the optimizer.","pith_inferences":["If the composition approximation holds, the same recipe should transfer to other differentiable constraint energies, such as no-fly zones, plume-impingement limits, or communication line-of-sight constraints, since the framework only requires a differentiable energy layer.","The paper leaves open whether a small number of MCMC refinement steps would close the gap between the approximate reverse-diffusion samples and the exact composed distribution; a testable extension is to measure constraint violation and sample diversity as a function of the number of MCMC steps.","The hand-tuned weights ($\\alpha_1=1.3$, $\\alpha_2=0.3$, point weights 1 and 2) suggest the method's robustness should be checked against variations in these constants, since a constraint-satisfaction result that survives weight perturbation would be much stronger than one tuned per scenario.","The reported distributional bias of the base diffusion model (e.g., negative-x bias up to 0.7 and positive-y bias less than 1.8) implies that composed constraints may inherit or amplify base-model bias; an extension would be to measure landing-site selection bias against the risk map rather than only aggregate violation rates."],"forward_implications":["A single trajectory diffusion model can be reused for new landing-site maps and constraints at inference time, eliminating the need to regenerate datasets and retrain for each mission variation.","Negation composition with a differentiable glideslope energy cuts glideslope violations from about 10% of segments to below 5%.","Inpainting enforces boundary and waypoint equality constraints exactly while preserving multi-modal trajectory choices, such as passing left or right of a waypoint.","Composing with a normalized landing-risk map reduces landings in regions with over 10% risk from 11% to under 7%.","Generated samples act as initial guesses for SCvx-based powered descent guidance, yielding dynamically feasible trajectories in a fraction of the time needed to generate optimizer solutions (about 1.17 minutes for 1,000 diffusion samples versus roughly one hour for 1,000 optimizer trajectories)."],"supporting_citations":[{"why":"Supplies the compositional diffusion formulation, the score-error bound, and the single-landing-site TrajDiffuser model from which this paper's model is initialized.","marker":"[14]"},{"why":"Establishes the product, mixture, and negation composition rules and the independence assumption underlying Eq. (16).","marker":"[13]"},{"why":"Provides the energy-based diffusion formulation, the score-gradient relationship, and the ULA-MCMC interpretation of reverse diffusion.","marker":"[32]"},{"why":"Introduces trajectory-level diffusion planning and inpainting for enforcing state and control constraints.","marker":"[12]"},{"why":"Supports the claim that diffusion policies represent multi-modal action distributions, motivating the multi-landing-site and waypoint results.","marker":"[17]"},{"why":"Defines the 6-DoF powered descent guidance problem and the SCvx parameters used to generate the training dataset.","marker":"[34]"},{"why":"Provides the discretized SOCP subproblem solved iteratively to create the trajectory dataset.","marker":"[35]"},{"why":"Formulates the multi-landing-site selection problem that motivates the risk-map composition.","marker":"[33]"}],"fun_headline_variants":["Diffusion composes constraints for safe landings","Compose diffusion PDFs to pick safe landing sites at runtime","Multi-modal descent guesses from composed diffusion models","Risk-map diffusion cuts unsafe landings by 40%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes that composing independently trained diffusion models by simple energy addition and negation, with only reverse diffusion acting as sampling, yields samples close enough to the true constrained distribution that the new constraints are satisfied.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion composes constraints for safe landings","Compose diffusion PDFs to pick safe landing sites at runtime","Multi-modal descent guesses from composed diffusion models","Risk-map diffusion cuts unsafe landings by 40%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000822,"raw_usage":{"total_tokens":3621,"prompt_tokens":991,"completion_tokens":2630,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":2567}},"tokens_in":607,"tokens_out":2630,"duration_ms":20735,"temperature":1.0,"reasoning_tokens":2567,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:39:45.952855+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the glideslope composition experiment and measure the fraction of trajectory segments violating the constraint for the base and composed models; if the composed model does not reproducibly cut the violation rate below the base model's roughly 10%, or if the reduction disappears when the penalty weight $\\lambda_{\\text{penalty}}$ is perturbed, the composition assumption is not doing the claimed work.","supporting_citations":[{"cited_title":"Compositional Diffusion Models for Powered Descent Trajectory Generation with Flexible Constraints,","cited_arxiv_id":null,"evidence_quote":"Supplies the compositional diffusion formulation, the score-error bound, and the single-landing-site TrajDiffuser model from which this paper's model is initialized."},{"cited_title":"Compositional Generative Modeling: A Single Model is Not All You Need,","cited_arxiv_id":null,"evidence_quote":"Establishes the product, mixture, and negation composition rules and the independence assumption underlying Eq. (16)."},{"cited_title":"Reduce, Reuse, Recycle: Compositional Generation with Energy-Based Diffusion Models and MCMC,","cited_arxiv_id":null,"evidence_quote":"Provides the energy-based diffusion formulation, the score-gradient relationship, and the ULA-MCMC interpretation of reverse diffusion."},{"cited_title":"Planning with Diffusion for Flexible Behavior Synthesis,","cited_arxiv_id":null,"evidence_quote":"Introduces trajectory-level diffusion planning and inpainting for enforcing state and control constraints."},{"cited_title":"Diffusion Policy: Visuomotor Policy Learning via Action Diffusion,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that diffusion policies represent multi-modal action distributions, motivating the multi-landing-site and waypoint results."},{"cited_title":"Successive Convexification for 6-DoF Mars Rocket Powered Landing with Free-Final-Time,","cited_arxiv_id":null,"evidence_quote":"Defines the 6-DoF powered descent guidance problem and the SCvx parameters used to generate the training dataset."},{"cited_title":"HALO: Hazard-Aware Landing Optimization for Autonomous Systems,","cited_arxiv_id":null,"evidence_quote":"Formulates the multi-landing-site selection problem that motivates the risk-map composition."}],"review_version":1}