{"id":"5e5c11e4-294c-4f74-98b1-141523c15dce","arxiv_id":"2412.09342","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A diffusion policy can satisfy novel state and action constraints by projecting each denoising step onto a dynamics-constrained feasible set and tightening those constraints to account for model mismatch.","lead":"This paper presents a way to make diffusion-based robot policies obey new safety constraints at test time, even when those constraints were absent from the training data. It repeatedly projects generated trajectories onto a dynamics-aware feasible set during denoising, and it tightens the constraints to absorb model error.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's closed-loop guarantee presupposes recursive feasibility of the tightened projection problem, which is neither stated nor proven; states in S_t but outside the tightened set may admit no feasible trajectory into \\tilde Z_f.","rationale":"The paper's central contribution is the claim that incorporating model-based projections into the backward diffusion process yields trajectories 'guaranteed to satisfy constraints.' The proof of that guarantee is Theorem 2, and its weakest link is the implicit assumption that the tightened feasible set \\tilde Z_f remains nonempty and reachable at every replanning step. The reader's weakest_assumption identifies exactly this: the induction establishes only st ∈ S_t, whereas the sampling step requires either st ∈ \\tilde S_t or a separate proof that some feasible trajectory into the tightened future sets exists. My reading of the manuscript in full does not reveal any other assumption that is more load-bearing: the nonconvexity of the experimental constraint sets and the use of a local SLSQP solver are additional reasons the empirical guarantee falls short of 1.0, but even in a purely convex linear setting the recursive-feasibility problem remains. This concern does not invalidate the empirical results, which are well supported by the reported ablations, baseline comparisons, and public code; it does mean the safety guarantee is overstated as written. Since the reader already arrived at CONDITIONAL on this basis, my stress-test leaves the verdict unchanged.","tokens_in":13057,"tokens_out":8621,"duration_ms":92348,"concrete_test":"Analytic check: instantiate Theorem 2 in a minimal convex setting with S=[0,1], A=[-0.05,0.05], dynamics s_{k+1}=s_k+a_k, γ=0.1, H=2, tightened sets \\tilde S_k=[0.1,0.9], and fixed current state s_0=0. Solve projection (9) onto \\tilde Z_f exactly. If the QP (or SLSQP, as the paper uses) reports infeasibility, then a state in S_0 but not in \\tilde S_0 makes the theorem's sampling premise impossible, showing the recursive-feasibility gap is real even when Theorem 1's convexity assumptions hold. For a direct check on the paper's setup, run DPCC with initial states placed on the boundary of the circular or halfspace constraints and record whether any projection call returns infeasible; even one infeasible replanning step falsifies the 'guaranteed' wording.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DPCC's headline safety claim rests on Theorem 2, whose induction step requires that at every replanning time t the controller can sample a trajectory from the tightened feasible set \\tilde Z_f. The proof only establishes st+1 ∈ S_{t+1} from st ∈ S_t; it never establishes st+1 ∈ \\tilde S_{t+1}, nor does it prove that the tightened projection problem is feasible at the next step. As Theorem 2 states \\tilde Z_f, membership requires st ∈ \\tilde S_t (the trajectory includes the current state), so the sampling premise can fail when the actual state lies in S_t but not in \\tilde S_t. If the intended reading is that only future predicted states are tightened, the theorem still needs a reachability or invariance argument showing that from every st ∈ S_t some feasible trajectory into \\tilde S_{t+1:...} exists; no such argument is supplied. Thus the deterministic closed-loop guarantee is conditional on an unproven recursive-feasibility assumption. The empirical numbers in Table 1 (constraints-and-goal rate 0.98, nonzero constraint-violation standard deviation) are consistent with this gap. This is fixable by stating and proving a recursive-feasibility or terminal-set condition, or by softening the word 'guaranteed'; the empirical contribution is not undermined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DPCC, an inference-time algorithm that combines an offline-trained trajectory diffusion policy with receding-horizon control and model-based projections onto state, action, and dynamics constraint sets. During each denoising step, candidate trajectories are projected into a feasible set, and the constraints are tightened by a model-mismatch bound. Two batch-selection criteria (temporal consistency and cumulative projection cost) are introduced, and the method is evaluated in a simulated 2D robot manipulator task with novel circular and halfspace obstacles. The central claims are that DPCC generates constraint-satisfying, dynamically feasible, goal-reaching trajectories, and that the constraint-tightening mechanism provides a closed-loop guarantee against model mismatch.","tokens_in":13300,"tokens_out":5262,"duration_ms":56435,"significance":"If the formal claims can be made correct, DPCC is a practically useful contribution: it addresses an important problem (enforcing novel test-time constraints for offline diffusion policies without retraining), the idea of iterating model-based projections inside the backward diffusion process is elegant, and the experimental comparison against guidance, post-processing, and model-free projection baselines is informative. The paper ships code and reports computation times and hyperparameters, which strengthens reproducibility. The relative empirical conclusions are credible; the main weakness is that the theoretical guarantee is stronger than what is proven.","major_comments":[{"comment":"The closed-loop guarantee is conditional on an unproven recursive-feasibility assumption. The theorem's premise states that at each timestep t one can sample a trajectory from p_theta(·|c, tilde Z_f), which requires the current state to lie in the tightened set tilde S_t, since the first predicted state st|t is constrained by tilde Z_f. The induction step, however, only proves st+1 in S_{t+1} from st in S_t; it never proves st+1 in tilde S_{t+1}, and Eq. (17) defines tilde S_{t+1} but does not specify the base case tilde S_0. Consequently, for a state in S_t \\ tilde S_t the problem may be infeasible and the premise can fail. A recursive-feasibility or terminal-invariance argument is needed, or the word \"guaranteed\" in the contributions and abstract should be softened.","section":"Section 5.3, Theorem 2"},{"comment":"The deterministic constraint-satisfaction claim in Eq. (16) and Theorem 2 presupposes that the projection onto Z_f (resp. tilde Z_f) is computed exactly. In the experiments, however, the state constraint sets are explicitly non-convex (St = {As <= b, ||s-p||_2^2 >= r^2}) and the projections are solved with the local SLSQP solver. A local solution of a non-convex QCQP need not be the global projection, so the iterates tau^{k-1} in Algorithm 1 are not guaranteed to lie in tilde Z_f. This is not a fatal flaw for the empirical comparison, but it invalidates the formal guarantee in the experimental setting; the gap should be acknowledged and the theoretical statements qualified.","section":"Section 6.1 and Appendix B"},{"comment":"Eq. (16) is not the sampling update derived in Theorem 1. Theorem 1 justifies the update tau^{k-1} ~ N(Pi_Zf(mu_theta), sigma_k^2 I), whereas Eq. (16) projects mu_theta + sigma_k epsilon_k after adding noise. These two operations produce different distributions, and the paper does not prove that iterating Eq. (16) approximates the conditional distribution p_theta(tau | O=1). Since the task-performance argument relies on preserving the learned trajectory distribution, this discrepancy needs either a formal justification or an explicit statement that Eq. (16) is a heuristic modification whose distributional properties are not guaranteed.","section":"Section 5.2, Theorem 1 and Eq. (16)"}],"minor_comments":[{"comment":"The sentence defining the action constraints is duplicated: \"The action constraints are defined as At = A, where A is the smallest bounding box containing all actions from the demonstration dataset (2)\" appears twice in the same paragraph.","section":"Section 6.1"},{"comment":"There is a capitalization and wording error after Eq. (16): \"By using (16), We denote the samples projected distribution as tau^0 ~ p_theta(·|c, Z_f)\" should be \"we denote\" and the syntax should be corrected.","section":"Section 5.2"},{"comment":"The notation for tightened feasible sets is inconsistent: Eq. (16) uses Z_f, while Algorithm 1 and Theorem 2 use tilde Z_f. The relation between the untightened and tightened projections should be made explicit at the point where the algorithm first introduces tilde Z_f.","section":"Section 5.3 and Algorithm 1"},{"comment":"The model-mismatch bound gamma is estimated from 100 unconstrained rollouts of the same policy and then used to tighten the constraints. As reported, this is a heuristic estimate rather than a certified bound; Theorem 2 relies on the assumption ||w_t||_2 <= gamma, so the practical safety claim inherits the uncertainty of this estimate. The authors should state this limitation explicitly.","section":"Section 6.1"},{"comment":"The conclusion states that DPCC can handle time-varying constraints directly, but the experiments do not evaluate time-varying constraints and no formal argument is given. Either add an experiment or soften the claim.","section":"Section 7"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution and the code release are valuable, and I see no reason to doubt the relative comparisons. The main issue is that the strong formal language used in the contributions, abstract, and Theorem 2 exceeds what is actually proven: recursive feasibility of the tightened problem is assumed, and the non-convex numerical projections in the experiments do not satisfy the exactness assumptions. These gaps are fixable by adding a recursive-feasibility/terminal-set analysis or by rephrasing the guarantee as a heuristic property, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"DPCC adds three things to the diffusion-policy toolbox: projecting the denoised trajectory onto a dynamics-aware feasible set at every diffusion step, tightening state constraints to absorb a bounded model mismatch, and two trajectory-selection heuristics (temporal consistency, cumulative projection cost). The experiments in the Avoiding benchmark are competently done, with public code and ablations that isolate each contribution. The gains over cost guidance and model-free projection are real. This is a solid incremental-technical contribution, not a field-shaping one.\n\nThe soft spots are in the formal claims. Theorem 2's 'guarantee' presupposes recursive feasibility. The proof begins with s_t in S_t, then assumes you can sample a trajectory from the tightened set \\tilde Z_f, which requires s_t in \\tilde S_t. The induction only proves s_{t+1} in S_{t+1}, not that the next replanning problem is feasible from the tightened set. So the closed-loop guarantee is conditional on an unproven invariance or reachability assumption. This is the standard recursive-feasibility issue in MPC, and it is fixable either by proving it under a terminal invariant set or by softening 'guaranteed' to 'empirically high-probability'.\n\nTwo smaller issues: Eq. (16) gives deterministic constraint satisfaction only for exact projections onto convex Z_f, but the experiments use nonconvex obstacle sets solved with SLSQP, so the projection is only local and Table 1's 0.98 rate is the honest number. And the bound gamma on model mismatch is estimated from rollouts, not certified, so the tightening is heuristic in practice.\n\nNone of this kills the paper. The algorithm works well in the tested setting, and the iterative-projection-plus-tightening recipe is worth having in the literature. What needs to change: clarify the recursive-feasibility assumption and the nonconvexity caveat, and tone down 'guaranteed' in the abstract and contributions. I'd send it to review, expecting minor-to-moderate revisions. For your own work, if you build on it, treat the guarantee as a heuristic and keep the safety analysis separate.","headline":"DPCC is a practical, well-evaluated method for constraining diffusion policies at test time, but its headline guarantee rests on an unproven recursive-feasibility assumption.","tokens_in":13836,"tokens_out":4763,"would_cite":true,"duration_ms":42624,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Diffusion policies trained offline can be made to satisfy novel test-time constraints by inserting a dynamics-aware projection into every denoising step and tightening constraints by the known model-error bound.","keywords":["diffusion policies","trajectory diffusion","model predictive control","constraint satisfaction","imitation learning","receding horizon control","constraint tightening","robot manipulation"],"falsifier":"Run the closed-loop experiment under the assumed model-mismatch bound $\\gamma$ and, at every replanning step, record whether the current state lies in the tightened set $\\tilde S_t$. If a state in $S_t$ but outside $\\tilde S_t$ is ever used as the start of a plan, the premise of Theorem 2 is void; and if constraints are violated while $\\|w_t\\|_2 \\le \\gamma$ holds, the claimed guarantee fails. A single such episode would settle the matter.","tokens_in":12785,"feed_emoji":"🤖","tokens_out":13488,"duration_ms":112077,"temperature":0.7,"pith_summary":"The paper claims that an offline-trained trajectory diffusion policy can be turned into a controller that respects novel state and action constraints absent from its training data, without retraining. The device is to modify each backward-diffusion step so the sampled trajectory is projected onto a set of dynamically feasible, constraint-satisfying trajectories, and to tighten the constraints by a margin that absorbs the bounded mismatch between the model used for projection and the true system. In the robot-manipulator simulations, this combination raises the rate of reaching the goal while satisfying constraints from below 50% (projection without tightening) to approximately 98%, with essentially zero constraint violations. If the claim is right, diffusion-based imitation learning gains the online constraint-handling ability of model predictive control while keeping the expressiveness of generative models.","feed_headline":"Diffusion policies obey new constraints without retraining","feed_subtitle":"Model-based projections plus constraint tightening reach ~98% success on novel obstacles while preserving goal-reaching.","key_machinery":"The operative object is the constraint- and model-based projection $\\Pi_{Z_f}$, defined as the nearest trajectory (in $\\ell^2$) that satisfies the state and action constraints and the dynamics equality constraints. It is applied to the noisy denoising output at each of the $K$ diffusion steps, which makes the final trajectory feasible by construction. The second piece is the tightening operation $\\tilde S_{t+1} = S_{t+1} \\ominus B_\\gamma$, which inflates the safety margin by the worst-case disturbance $\\gamma$; this absorbs model mismatch so that predicted feasibility implies actual feasibility under the true dynamics. Together they convert a learned generative prior into a receding-horizon controller whose closed-loop trajectories are dynamically feasible and constraint-satisfying.","core_discovery":"The central claim is that constraint satisfaction can be guaranteed, not merely encouraged, by embedding model-based projections into the denoising process. The paper defines the feasible trajectory set $Z_f$ of trajectories whose states and actions satisfy the constraints and whose dynamics agree with the model $f$, and replaces the plain denoising step with the model-informed step $\\tau^{k-1} = \\Pi_{Z_f}(\\mu_\\theta(\\tau^k,k,c) + \\sigma_k \\epsilon_k)$ (Eq. 16). The projection is applied at every diffusion iteration, so the final sample lies in $Z_f$ for any noise schedule. To handle the discrepancy between $f$ and the true dynamics $s_{t+1}=f(s_t,a_t)+w_t$ with $\\|w_t\\|_2 \\le \\gamma$, the state constraints are tightened to $\\tilde S_{t+1} = S_{t+1} \\ominus B_\\gamma$; Theorem 2 then asserts that if each replanning samples from the tightened feasible set, all future actual states stay inside the original constraints. The paper also supplies two trajectory-selection criteria to avoid mode-switching and to preserve the learned task behavior, and reports that the best variant, DPCC-C, reaches the goal with constraint satisfaction in 98% of episodes while taking fewer timesteps than the baselines.","pith_inferences":["Because the projection is redefined at each replanning step, the same algorithm should extend directly to time-varying constraints such as moving obstacles; the paper notes this but does not test it, and a moving-obstacle benchmark would be a natural stress test.","The experiments solve non-convex projected sets with a nonlinear optimizer even though Theorem 1 assumes convexity, so the convexity assumption appears conservative; analyzing the non-convex projection's effect on the sampling distribution would either justify or refine the method.","The cumulative projection-cost selection criterion can be read as a proxy for staying close to the learned data distribution, hinting at a principled link between projection-guided sampling and likelihood-preserving guidance.","In practice the guarantee rests on knowing a valid mismatch bound $\\gamma$; where that bound is uncertain, an online estimator or an outer safety filter would be needed to preserve the near-100% constraint satisfaction rate."],"forward_implications":["An offline-trained diffusion policy can be deployed in environments with constraints never seen in training, simply by changing the projection sets at test time.","Because the projection enforces the dynamics model at every denoising step, the generated trajectories remain executable by the low-level controller, unlike constraints that are only projected once after sampling.","With a known upper bound on model mismatch, the constraint-tightening argument turns predicted feasibility into actual feasibility, which is why adding tightening raises the combined goal-and-constraint success rate from below 50% to about 98%.","The two selection criteria, temporal consistency and cumulative projection cost, reduce mode-switching and preserve goal-reaching performance, with the cumulative-cost criterion also reaching the goal fastest.","Each control action costs about 80 ms of computation, so DPCC is usable at moderate replanning rates on a workstation GPU but not at high control frequencies."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model whose forward and backward processes the trajectory sampler is built on.","marker":"Ho et al., 2020"},{"why":"Introduces the trajectory-diffusion receding-horizon framework that DPCC follows to generate and replan trajectories.","marker":"Janner et al., 2022"},{"why":"Provides the classifier-guidance derivation, based on a Taylor expansion of the log-likelihood, used in the proof of Theorem 1.","marker":"Dhariwal and Nichol, 2021"},{"why":"Supplies the control-as-inference formulation with the binary feasibility variable O that defines the conditional sampling target.","marker":"Toussaint, 2009"},{"why":"Gives existence and uniqueness of the projection onto a closed convex set, invoked in the proof of Theorem 1.","marker":"Bazaraa et al., 2006"},{"why":"Defines the model-free projection baseline that DPCC extends with dynamics awareness and outperforms in the simulations.","marker":"Römer et al., 2024"},{"why":"Provides the gradient-guidance baseline that encodes constraints as cost gradients during denoising, one of the approaches DPCC is compared against.","marker":"Carvalho et al., 2023"},{"why":"Another gradient-guidance baseline for constrained diffusion policies used in the experimental comparison.","marker":"Kondo et al., 2024"}],"fun_headline_variants":["Projection-based diffusion guarantees novel constraints","Diffusion control with provable novel obstacle handling","DPCC: enforce new constraints via denoising projections","Diffusion policies gain constraint guarantees via projection","Model-based projections let diffusion adapt to new limits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee depends on being able to keep sampling trajectories that satisfy the stricter, safety-margin constraints at every replanning step, which requires the robot to remain inside that stricter region; the proof only shows it remains inside the original, looser region, so that requirement is assumed rather than proven.","fun_headline_variants_meta":{"raw":{"variants":["Projection-based diffusion guarantees novel constraints","Diffusion control with provable novel obstacle handling","DPCC: enforce new constraints via denoising projections","Diffusion policies gain constraint guarantees via projection","Model-based projections let diffusion adapt to new limits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00073,"raw_usage":{"total_tokens":3274,"prompt_tokens":957,"completion_tokens":2317,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2247}},"tokens_in":573,"tokens_out":2317,"duration_ms":17868,"temperature":1.0,"reasoning_tokens":2247,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:06:04.798305+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the closed-loop experiment under the assumed model-mismatch bound $\\gamma$ and, at every replanning step, record whether the current state lies in the tightened set $\\tilde S_t$. If a state in $S_t$ but outside $\\tilde S_t$ is ever used as the start of a plan, the premise of Theorem 2 is void; and if constraints are violated while $\\|w_t\\|_2 \\le \\gamma$ holds, the claimed guarantee fails. A single such episode would settle the matter.","supporting_citations":[{"cited_title":"Planning with D iffusion for F lexible B ehavior S ynthesis","cited_arxiv_id":null,"evidence_quote":"Introduces the trajectory-diffusion receding-horizon framework that DPCC follows to generate and replan trajectories."},{"cited_title":"Robot T rajectory O ptimization using A pproximate I nference","cited_arxiv_id":null,"evidence_quote":"Supplies the control-as-inference formulation with the binary feasibility variable O that defines the conditional sampling target."},{"cited_title":"Nonlinear P rogramming: T heory and A lgorithms","cited_arxiv_id":null,"evidence_quote":"Gives existence and uniqueness of the projection onto a closed convex set, invoked in the proof of Theorem 1."}],"review_version":1}