betaFeedback

WorksComputer Science

Test-Time Constraint-Aware Diffusion Predictive Control

Abstract

Diffusion models can learn multimodal distributions over finite-horizon state-input sequences, providing a way to generate candidates for predictive control. However, sequences generated by a pretrained model may violate safety constraints introduced at deployment, for example when obstacles are added or repositioned. Retraining whenever these constraints change limits the usefulness of learned predictive controllers. In this paper, we propose Test-Time Constraint-Aware Diffusion Predictive Control (TCA-DPC), an algorithm for adapting the sampling distribution of a pretrained diffusion controller to newly introduced safety constraints without retraining. We treat the pretrained model as a prior over finite-horizon sequences and define an adapted distribution through exponential tilting based on a constraint-violation penalty. The tilt strength is adjusted online using the fraction of generated sequences that fail constraint verification. We approximately sample the adapted distribution using the Metropolis-adjusted Langevin algorithm in the Gaussian noise space of the pretrained diffusion sampler.

Introduction

In predictive control, multiple input sequences may achieve the same objective while exhibiting different behaviours. For example, different initializations of a nonlinear model predictive control solver may yield distinct locally optimal solutions [10, 13], while human demonstrations may follow different routes around an obstacle. Such variations can give rise to multimodal distributions over control sequences or joint state-input trajectories. Diffusion models provide a way to learn these distributions from data obtained through numerical optimization, simulation, or human demonstrations [11, 23]. Once trained, the model generates finite-horizon sequences through a reverse denoising process, allowing different modes of the learned distribution to be sampled. In what follows, we refer to the trained diffusion model together with its reverse denoising procedure as the pretrained diffusion sampler.

The distribution induced by the pretrained sampler reflects the behaviours and constraints represented in the training data. At deployment, additional safety requirements may arise, for example when a new obstacle occupies a previously traversable region. Sequences that remain likely under the pretrained sampler may then violate the deployment constraints. Retraining the diffusion model whenever these requirements change is costly. We therefore seek to adapt the pretrained sampler to the deployment constraints while keeping the trained model parameters fixed.

Existing approaches modify diffusion sampling through guidance [1, 8] or projection [7]. In robot motion planning, Motion Planning Diffusion uses gradient guidance to incorporate planning objectives and previously unseen obstacles [6]. For predictive control, Römer et al. [19] propose Diffusion Predictive Control with Constraints (DPCC), which incorporates model-based projections into the DDPM denoising process [9]. Their experiments demonstrate improved constraint satisfaction and task performance over the guidance and projection baselines considered. DPCC formulates these projections using an explicit description of the state and input constraints, including obstacle geometry in its navigation experiments. It also tightens the state constraints using a model-mismatch bound, effectively adding clearance around obstacles. Its robustness guarantee depends on the validity of this bound. These requirements motivate an alternative that accesses deployment constraints through an evaluator providing violation penalties and their derivatives, without requiring an explicit geometric description of the feasible set, a projection operator, or a prescribed model-mismatch bound.

In this paper, we develop Test-Time Constraint-Aware Diffusion Predictive Control (TCA-DPC), a gradient-based procedure for adapting a pretrained diffusion sampler to new safety constraints in predictive control. We take the distribution induced by the pretrained sampler as the pretrained diffusion prior and reweight it through a change of measure via an exponential weight, also known as exponential tilting, which admits an interpretation as loss-based updating of the prior distribution [2]. The exponential weight depends on a penalty for violating the deployment constraints, so sequences with larger penalties receive smaller weights. All sequences with zero penalty receive the same weight before normalization. Consequently, the tilted target preserves the relative probabilities of feasible behaviours whenever the prior assigns positive probability to feasible sequences. We implement this adaptation by changing the distribution of the random inputs supplied to the pretrained sampler. The trained denoising network remains unchanged. This allows us to use derivatives of the violation penalty through the complete denoising process to construct proposals in the noise domain. We also adjust the tilt strength during control using feedback from the verification of generated sequences. Our contributions are as follows:

  • We construct a target distribution in noise space whose image under the pretrained diffusion sampler is the desired exponentially tilted distribution over sequences. We approximately sample this target using the Metropolis-adjusted Langevin algorithm (MALA) [18], with proposals based on gradients through the complete sampler.

  • We introduce an online feedback rule for the tilt strength. At each control step, the rule compares the observed fraction of generated sequences that fail verification with a prescribed target failure rate. It increases or decreases the tilt strength accordingly, within a specified range, without updating the pretrained diffusion model.

Related work. Exponential tilting has also been used to adapt diffusion-generated distributions in other settings. DiffTilt [14] reweights a learned joint distribution over environments and system executions to search for safety-critical failures. Our objective is instead to generate sequences with fewer violations of deployment constraints. BayesFP [21] formulates inference-time adaptation of pretrained diffusion and flow models as posterior sampling with a cost-derived likelihood. It builds on the Feynman-Kac corrector framework [22] to approximate the resulting distribution. Concurrent work, MOSAIC [3], also adapts a distribution over the stochastic inputs of a pretrained diffusion model and adjusts constraint penalties using feedback. It uses gradient-free, importance-weighted updates of mode-specific proposal distributions. Our approach uses derivatives of the violation penalty through the complete diffusion sampler to construct MALA proposals in noise space. The feedback also differs: MOSAIC updates its constraint multipliers using importance-weighted estimates of constraint costs, whereas our update uses the observed fraction of sequences that fail verification across successive control steps.

Problem Setup

Consider the discrete-time dynamical system

st+1=f(st,ut),f:Rn×Rm→Rn,(1)s_{t+1} = f(s_t, u_t), \qquad f : \mathbb{R}^{n} \times\mathbb{R}^{m} \to\mathbb{R}^{n}, \tag*{(1)}

where sts_t is the state and utu_t is the applied control input. The state-transition function ff is time-invariant and need not be known exactly. The control objective is to steer the system toward a prescribed target while satisfying state and input constraints, including safety requirements introduced at deployment. In conventional model predictive control [5], this objective is addressed by planning, at each control time tt, future inputs and states over a prediction horizon of NN stages through a finite-horizon optimization using an approximate dynamics model f^\hat{f}.

Here, a pretrained diffusion model generates candidate state-input sequences for this planning step. In particular, a diffusion model is trained on finite-horizon segments of offline expert demonstrations (similar to imitation learning [17]), it approximates their conditional distribution, which can represent different ways of reaching the target. At control time tt, a generated sequence takes the form

Y=((s^0∣t,u^0∣t),…,(s^N−1∣t,u^N−1∣t))∈Y⊆Rd,(2)Y = \left((\hat{s}_{0|t}, \hat{u}_{0|t}), \ldots, (\hat{s}_{N-1|t}, \hat{u}_{N-1|t})\right) \in\mathcal{Y} \subseteq\mathbb{R}^{d}, \tag*{(2)}

where d=N(n+m)d = N(n+m). The quantities s^k∣t\hat{s}_{k|t} and u^k∣t\hat{u}_{k|t} denote the generated state prediction and planned input for time t+kt+k respectively, based on the information available at time tt, with s^0∣t=st\hat{s}_{0|t} = s_t.

Let ctc_t denote the conditioning information, including the current state and any task information used by the sampler. For example, in a navigation task, this may specify a goal line, with the episode ending when the system reaches it. For fixed ctc_t, the pretrained sampler, driven by fresh Gaussian random inputs, generates

Y∼Pt,Pt(⋅):=Pθ(⋅∣ct),(3)Y \sim\mathbb{P}_t, \qquad\mathbb{P}_t(\cdot) := \mathbb{P}_{\theta}(\cdot\mid c_t), \tag*{(3)}

where θ\theta denotes the trained parameters. We call Pt\mathbb{P}_t the pretrained diffusion prior.

The controller operates in a receding-horizon manner: at each control step, it selects a sequence from the generated candidates and applies only its first input, ut=u^0∣tu_t = \hat{u}_{0|t}. After observing st+1s_{t+1}, it generates new sequences and repeats the procedure until the episode’s termination condition is met. The episode length TT counts the executed control steps and may vary between episodes; it need not equal the prediction horizon NN. An individual generated sequence therefore need not reach the target within its prediction horizon. Our objective is to adapt the sampling distribution to the deployment constraints while keeping the trained parameters θ\theta of the diffusion model fixed.

Diffusion model. We use a pretrained denoising diffusion probabilistic model (DDPM) Ho et al. [9]. During training, sequences are corrupted with Gaussian noise, and a neural network learns to predict the added noise. Generation proceeds through a learned reverse denoising process. Let Y(i)Y^{(i)} denote the sequence at diffusion step ii, with KK denoising steps and Y(0)Y^{(0)} the generated sequence. Sampling starts from Y(K)∼N(0,Id)Y^{(K)} \sim\mathcal{N}(0, I_d) before conditioning. Each reverse update, before reimposing the conditioning values, takes the form Y(i−1)←μθ,i(Y(i))+σiξiY^{(i-1)} \leftarrow\mu_{\theta,i}(Y^{(i)}) + \sigma_i \xi_i, i=K,…,1i = K, \ldots, 1, where μθ,i\mu_{\theta,i} is the DDPM reverse mean computed from the noise-prediction network and the prescribed noise schedule Janner et al. [11]. The variables ξi∼N(0,Id)\xi_i \sim\mathcal{N}(0, I_d) are independent of one another and of Y(K)Y^{(K)}. The sampling noise scales σi\sigma_i are prescribed, with σ1=0\sigma_1 = 0. In this paper, conditioning on the current state is enforced by fixing the corresponding state entries at initialization and after every reverse update. These operations are determined by ctc_t. The resulting sequence Y=Y(0)Y = Y^{(0)} has distribution Pθ(⋅∣ct)\mathbb{P}_{\theta}(\cdot\mid c_t).

Test-time safety constraints. The pretrained diffusion prior reflects the behaviours and constraints represented in the training data. At deployment, additional requirements may arise, for example when an obstacle occupies a previously traversable region. These requirements can exclude sequences that remain likely under the pretrained model. We express the requirements through signed constraint residuals

gℓ,t(Y)≤0,ℓ=1,…,e,(4)g_{\ell,t}(Y) \le0,\qquad\ell= 1,\ldots,e, \tag*{(4)}

where e∈Ne \in\mathbb{N}, and gℓ,t:Y→Rg_{\ell,t} : \mathcal{Y} \to\mathbb{R}. The index ℓ\ell enumerates the constraints over all relevant prediction stages. Their evaluation depends on the quantities contained in YY and the prediction procedure used by the underlying controller. Dependence on the current state and other information in ctc_t is implicit in the index tt. The residuals may also include nominal admissibility and model-consistency requirements. Any constraint tightening is incorporated into their definition. The corresponding feasible set is At={Y∈Y:gℓ,t(Y)≤0 for every ℓ=1,…,e}\mathcal{A}_t = \{Y \in\mathcal{Y} : g_{\ell,t}(Y) \le0 \text{ for every } \ell= 1,\ldots,e\}. We assume access to a constraint evaluator that provides the violation penalty and the derivatives required by the sampling procedure. The adaptation does not require a projection operator onto At\mathcal{A}_t or an explicit description of its boundary. This distinction between oracle access and explicit constraint representations is standard in constrained convex optimization; see, for example, Lee et al. [15]; Mhammedi [16]. To quantify violations, let ρ:R→R≥0\rho: \mathbb{R} \to\mathbb{R}_{\ge0} be a continuously differentiable, non-decreasing function satisfying ρ(a)=0\rho(a) = 0 for a≤0a \le0, ρ(a)>0\rho(a) > 0 for a>0a > 0. Define

Φt(Y)=∑ℓ=1eρ(gℓ,t(Y)).(5)\Phi_t(Y) = \sum_{\ell=1}^{e} \rho(g_{\ell,t}(Y)). \tag*{(5)}

A suitable choice for ρ\rho is the squared hinge ρ(a)=[a]+2\rho(a) = [a]_+^2, where [a]+=max⁡{a,0}[a]_+ = \max\{a,0\}. This construction gives

Φt(Y)≥0,Φt(Y)=0⟺Y∈At.(6)\Phi_t(Y) \ge0,\qquad\Phi_t(Y) = 0 \quad\Longleftrightarrow\quad Y \in\mathcal{A}_t. \tag*{(6)}

We assume that Φt\Phi_t is measurable and finite Pt\mathbb{P}_t-almost surely; an upper bound on the penalty is not required.

Adaptation objective. Given the pretrained diffusion prior Pt\mathbb{P}_t and the deployment constraints, our objective is to construct and approximately sample an adapted distribution Qt\mathbb{Q}_t that increases the probability of constraint-satisfying sequences. The adaptation must operate within a prescribed sampling budget at each control step and use feedback from sequence verification without changing the trained parameters θ\theta. The following section constructs an exponentially tilted family Qβ,tQ_{\beta,t} and develops an online rule for adjusting its tilt strength β\beta.

Test-Time Constraint-Aware Diffusion Predictive Control

We adapt the pretrained diffusion prior by assigning smaller weights to sequences that violate the deployment constraints. The adaptation retains the relative prior probabilities of feasible sequences and leaves the trained parameters unchanged. We first define the adapted distribution, known as the diffusion posterior, then describe how to sample it approximately through the random inputs of the pretrained diffusion sampler. The adaptation strength, called the tilt parameter, is selected online using feedback from constraint verification. The proposed approach (TCA-DPC) combines three components: exponential tilting of the pretrained diffusion prior, approximate sampling through noise-space MALA, and an online feedback rule for adjusting the tilt strength.

Exponential tilting of the pretrained diffusion prior: The diffusion posterior

The diffusion posterior should depend on the safety violations associated with the generated sequences. We capture this dependence through a multiplicative factor that reweights the pretrained diffusion prior according to the violation penalty. To formalize this adaptation, we use a standard change-of-measure approach from probability theory, namely exponential tilting [2]. For a tilting parameter β>0\beta> 0, the diffusion posterior is defined as

Qβ,t(dY)=exp⁡[−βΦt(Y)]Zβ,tPt(dY),Zβ,t=EPt[exp⁡[−βΦt(Y)]].(7)Q_{\beta,t}(dY) = \frac{\exp[-\beta\Phi_t(Y)]}{Z_{\beta,t}}P_t(dY), \qquad Z_{\beta,t} = \mathbb{E}_{P_t}\left[\exp[-\beta\Phi_t(Y)]\right]. \tag*{(7)}

Since Φt\Phi_t is nonnegative and finite, 0<Zβ,t≤10 < Z_{\beta,t} \le1 for every finite β\beta. Increasing β\beta strengthens the preference for sequences with smaller violations, while Q0,t=PtQ_{0,t} = P_t. The purpose of Qβ,tQ_{\beta,t} is to reduce the probability of constraint violations while retaining the different feasible behaviours represented by the prior. The following proposition captures this notion.

Proposition 1 (Effect of exponential tilting) For a fixed control time tt with Pt(At)>0P_t(A_t) > 0, the tilted distributions defined in (7) satisfy Qβ,t(⋅∣At)=Pt(⋅∣At)Q_{\beta,t}(\cdot\mid A_t) = P_t(\cdot\mid A_t) for every finite β≥0\beta\ge0, while the violation probability Qβ,t(Atc)Q_{\beta,t}(A_t^c) is non-increasing in β\beta and converges to zero as β→∞\beta\to\infty.

The proof of Proposition 1 follows from straightforward arguments, but for completeness, we have added it to the supplementary material. Proposition 1 gives a precise interpretation of the adaptation: increasing β\beta reduces the violation probability without changing the relative probabilities of feasible sequences. These properties motivate adjusting β\beta online using feedback from constraint verification. Before updating β\beta online, we now describe how to draw samples from the diffusion posterior.

Sampling from the diffusion posterior

To sample from the diffusion posterior, we express the pretrained diffusion sampler as a deterministic function of its Gaussian random inputs. Collect the initial draw Y(K)Y^{(K)}, before conditioning, and the independent reverse-step noises into z=(Y(K),ξK,…,ξ2)∈Rdzz = (Y^{(K)}, \xi_K, \ldots, \xi_2) \in\mathbb{R}^{d_z}, \qquad z∼Γ:=N(0,Idz)z \sim\Gamma:= \mathcal{N}(0, I_{d_z}), omitting noises at steps with σi=0\sigma_i = 0. For fixed ctc_t, the complete denoising procedure, including conditioning and noise scaling, defines a map Ψct\Psi_{c_t} such that Y=Ψct(z)Y = \Psi_{c_t}(z). Thus, the pretrained diffusion prior Pt\mathbb{P}_t is the pushforward of Γ\Gamma under Ψct\Psi_{c_t}. For every measurable A⊆YA \subseteq\mathcal{Y}, the pushforward integration formula [4] gives

Qβ,t(A)=1Zβ,t∫Y1A(Y)e−βΦt(Y)Pt(dY)=1Zβ,t∫Ψct−1(A)e−βΦt(Ψct(z))Γ(dz),(8)\mathbb{Q}_{\beta,t}(A) = \frac{1}{\mathcal{Z}_{\beta,t}} \int_{\mathcal{Y}} 1_A(Y)e^{-\beta\Phi_t(Y)} \mathbb{P}_t(dY) = \frac{1}{\mathcal{Z}_{\beta,t}} \int_{\Psi_{c_t}^{-1}(A)} e^{-\beta\Phi_t(\Psi_{c_t}(z))} \Gamma(dz), \tag*{(8)}

where Ψct−1(A)={z:Ψct(z)∈A}\Psi_{c_t}^{-1}(A) = \{z : \Psi_{c_t}(z) \in A\}. Defining πβ,t(dz)=e−βΦt(Ψct(z))Zβ,tΓ(dz)\pi_{\beta,t}(dz) = \frac{e^{-\beta\Phi_t(\Psi_{c_t}(z))}}{\mathcal{Z}_{\beta,t}} \Gamma(dz), we obtain Qβ,t(A)=πβ,t(Ψct−1(A))\mathbb{Q}_{\beta,t}(A) = \pi_{\beta,t}(\Psi_{c_t}^{-1}(A)). Consequently, sampling z∼πβ,tz \sim\pi_{\beta,t} and applying Ψct\Psi_{c_t} produces a sequence with law Qβ,t\mathbb{Q}_{\beta,t}, without evaluating the density of Pt\mathbb{P}_t or requiring Ψct\Psi_{c_t} to be invertible. The density of πβ,t\pi_{\beta,t} is proportional to exp⁡[−Vβ,t(z)]\exp[-V_{\beta,t}(z)], with

Vβ,t(z)=12∥z∥2+βΦt(Ψct(z)).(9)V_{\beta,t}(z) = \frac{1}{2}\lVert z\rVert^2 + \beta\Phi_t(\Psi_{c_t}(z)). \tag*{(9)}

We use the Metropolis-adjusted Langevin algorithm (MALA) [18, 20] to approximately sample this target. Gradients are computed through the complete diffusion sampler and the violation penalty. Holding ctc_t and β\beta fixed, iteration j=0,…,M−1j = 0,\ldots,M-1 proposes

z′=z(j)−h∇zVβ,t(z(j))+2h ζj,ζj∼N(0,Idz),(10)z' = z^{(j)} - h\nabla_z V_{\beta,t}(z^{(j)}) + \sqrt{2h}\,\zeta_j,\qquad\zeta_j \sim\mathcal{N}(0,I_{d_z}), \tag*{(10)}

where h>0h > 0 is fixed and the ζj\zeta_j are independent. Let qβ,t(z′∣z)q_{\beta,t}(z' \mid z) denote the Gaussian proposal density with mean z−h∇zVβ,t(z)z - h\nabla_z V_{\beta,t}(z) and covariance 2hIdz2hI_{d_z}. The proposal is accepted with probability

aβ,t(z,z′)=min⁡{1,eVβ,t(z)−Vβ,t(z′)qβ,t(z∣z′)qβ,t(z′∣z)}.(11)a_{\beta,t}(z,z') = \min\left\{1,e^{V_{\beta,t}(z)-V_{\beta,t}(z')}\frac{q_{\beta,t}(z\mid z')}{q_{\beta,t}(z'\mid z)}\right\}. \tag*{(11)}

If accepted, z(j+1)=z′z^{(j+1)} = z'; otherwise, z(j+1)=z(j)z^{(j+1)} = z^{(j)}. The normalizing constant cancels from the acceptance ratio. For a fixed target, this Metropolis correction leaves πβ,t\pi_{\beta,t} invariant. After MM iterations, we generate Y=Ψct(z(M))Y = \Psi_{c_t}(z^{(M)}) as an approximate sample from Qβ,t\mathbb{Q}_{\beta,t}.

This construction provides an approximate sampler for Qβ,t\mathbb{Q}_{\beta,t}. It remains to select the tilt strength during deployment. Small values of β\beta leave the target distribution close to the pretrained diffusion prior, with Q0,t=Pt\mathbb{Q}_{0,t} = \mathbb{P}_t, and may provide insufficient preference for constraint-satisfying sequences. Larger values strengthen this preference but can make the target harder to explore. We therefore adjust the tilt strength online using feedback from sequence verification. At each control step tt, we hold βt\beta_t fixed while generating a batch of BB sequences and use the fraction that fail verification to determine βt+1\beta_{t+1}. The following subsection presents this feedback rule.

Online selection of the tilt parameter

We adjust the tilt strength using verification feedback, aiming to regulate the fraction of failing sequences around a prescribed maximum failure rate δ∈[0,1]\delta\in[0,1]. At control time tt, we hold βt\beta_t fixed and generate BB sequences {Yt(b)}b=1B\{Y_t^{(b)}\}_{b=1}^{B} using MALA targeting Qβt,t\mathbb{Q}_{\beta_t,t}. After the initial MM MALA steps, we compute the failure fraction and update the tilt for the next control step:

ft=1B∑b=1B1{Yt(b)∉At},(12)f_t = \frac{1}{B}\sum_{b=1}^{B}\mathbf{1}\left\{Y_t^{(b)} \notin\mathcal{A}_t\right\}, \tag*{(12)}
βt+1=clip⁡[0,βmax⁡](βt+κ(ft−δ)).(13)\beta_{t+1} = \operatorname{clip}_{[0,\beta_{\max}]}\left(\beta_t + \kappa(f_t-\delta)\right). \tag*{(13)}

where κ>0\kappa> 0 is the feedback gain, βmax⁡>0\beta_{\max} > 0 bounds the tilt strength, and clip⁡\operatorname{clip} projects its argument onto the indicated interval. We initialize β0∈[0,βmax⁡]\beta_0 \in[0,\beta_{\max}] and measure ftf_t before any additional refinement or fallback. The update implements projected integral feedback: it accumulates the discrepancy ft−δf_t - \delta, increasing the tilt above the maximum failure rate and decreasing it below the budget, subject to the bounds. This direction is motivated by Proposition 1: for a fixed context and constraint set, the violation probability under the ideal tilted distribution is non-increasing in β\beta. The cap βmax⁡\beta_{\max} is a computational design choice, since stronger tilting can make the target harder to explore within a finite MALA budget.

Receding-horizon implementation

The preceding subsections define the adapted sampler and its feedback update. For control execution, the controller selects a verified sequence minimizing an application-dependent cost JselJ_{\mathrm{sel}}, such as deviation from the previously selected sequence, and applies its first input. A finite sampling budget may produce no verified sequence and therefore, a fallback is necessary. Selecting a failed sequence at random ignores the severity of its predicted violations and could direct the system toward an unsafe region. We therefore first attempt a bounded number of additional MALA rounds at the current control step. Each round updates all BB noise vectors once, with βt\beta_t and hh fixed, and checks the regenerated sequences. Refinement stops when a sequence passes verification, and selection proceeds using JselJ_{\mathrm{sel}}.

If no sequence passes, the controller continues the previously selected sequence by applying its next planned input that has not yet been applied: for example, after applying u^0∣t−1\hat{u}_{0|t-1} at time t−1t - 1, reuse at time tt applies u^1∣t−1\hat{u}_{1|t-1}. If that sequence was verified, the constraints remain unchanged, and tracking is accurate, its remaining inputs provide a previously checked continuation. Reuse is permitted for a limited number of consecutive control steps, during which ordinary MALA refinement and tilt updates continue. This limit bounds reliance on earlier predictions. If no previous sequence is available or the reuse limit is reached, the controller applies the first input of the sequence with the smallest maximum constraint violation. These fallback choices are heuristic and do not guarantee eventual verification success.

Algorithm 1 summarizes the controller. The stored sequence YprevY_{\mathrm{prev}} contains the remaining state-input pairs of the most recently selected sequence; each pair is removed after its input is applied. Initially, all noise vectors are independent draws from N(0,Idz)\mathcal{N}(0,I_{d_z}), YprevY_{\mathrm{prev}} is empty, and nreuse=0n_{\mathrm{reuse}} = 0. This counter increases after each reuse and resets whenever a newly generated sequence is selected.

Numerical Study

We evaluate2 TCA-DPC ( Algorithm 1) on the Avoiding task from D3IL ([12]), using the test-time constraint setting of DPCC ([19]). Using the same pretrained models, we compare adaptive tilting with projection-based, untilted, and fixed-tilt baselines. We assess constraint violations and closed-loop performance, track the tilt strength throughout each episode, and examine sensitivity to constraint tightening.

Environment, data, and test-time constraints. In Avoiding, a simulated manipulator moves its end-effector to a goal line while avoiding six obstacles. Following DPCC ([19], Section 6.1), st∈R4s_t \in\mathbb{R}^{4} contains the current and desired planar end-effector positions, and ut∈R2u_t \in\mathbb{R}^{2} is the

Algorithm 1 TCA-DPC

Input: Context ctc_t; stored noise vectors {z(b)}b=1B\{z^{(b)}\}_{b=1}^{B}, tilt βt\beta_t, remaining previous sequence YprevY_{\mathrm{prev}}, and reuse counter nreusen_{\mathrm{reuse}}; JselJ_{\mathrm{sel}}, MM, hh, δ\delta, κ\kappa, βmax⁡\beta_{\max}; reuse limit Rmax⁡R_{\max} and additional refinement budget MextraM_{\mathrm{extra}}

// Sampling and feedback

Construct Φt\Phi_t and Vβt,tV_{\beta_t,t} using Eq. (9)

Refresh one z(b)∼N(0,Idz)z^{(b)} \sim\mathcal{N}(0,I_{d_z}); retain the other noise vectors

for b=1,…,Bb = 1,\ldots,B do

Perform MM MALA updates of z(b)z^{(b)} using Eqs. (10)–(11) with β=βt\beta= \beta_t and step size hh; Y(b)←Ψct(z(b))Y^{(b)} \leftarrow\Psi_{c_t}(z^{(b)})

end

I←{b:Y(b) passes verification}\mathcal{I} \leftarrow\{b:Y^{(b)}\text{ passes verification}\}; ft←1−∣I∣/Bf_t \leftarrow1-|\mathcal{I}|/B // measured before fallback

βt+1←clip⁡[0,βmax⁡](βt+κ(ft−δ))\beta_{t+1} \leftarrow\operatorname{clip}_{[0,\beta_{\max}]}(\beta_t+\kappa(f_t-\delta))

// Control selection and fallback

j←0j \leftarrow0 while I=∅\mathcal{I}=\varnothing and j<Mextraj<M_{\mathrm{extra}} do

Perform one further MALA update for each z(b)z^{(b)} with the same βt\beta_t and hh; regenerate Y(b)=Ψct(z(b))Y^{(b)}=\Psi_{c_t}(z^{(b)}), I←{b:Y(b) passes verification}\mathcal{I}\leftarrow\{b:Y^{(b)}\text{ passes verification}\};

j←j+1j\leftarrow j+1

end

if I≠∅\mathcal{I}\ne\varnothing then

b⋆←arg⁡min⁡b∈I Jsel(Y(b))b^\star\leftarrow\underset{b\in\mathcal{I}}{\arg\min}\ J_{\mathrm{sel}}(Y^{(b)}); Ysel←Y(b⋆)Y_{\mathrm{sel}}\leftarrow Y^{(b^\star)}; nreuse←0n_{\mathrm{reuse}}\leftarrow0

else if YprevY_{\mathrm{prev}} is nonempty and nreuse<Rmax⁡n_{\mathrm{reuse}}<R_{\max} then

Ysel←YprevY_{\mathrm{sel}}\leftarrow Y_{\mathrm{prev}}; nreuse←nreuse+1n_{\mathrm{reuse}}\leftarrow n_{\mathrm{reuse}}+1

else

b⋆←arg⁡min⁡b=1,…,B max⁡ℓ=1,…,e gℓ,t(Y(b))b^\star\leftarrow\underset{b=1,\ldots,B}{\arg\min}\ \underset{\ell=1,\ldots,e}{\max}\ g_{\ell,t}(Y^{(b)}); Ysel←Y(b⋆)Y_{\mathrm{sel}}\leftarrow Y^{(b^\star)}; nreuse←0n_{\mathrm{reuse}}\leftarrow0 // unverified selection

end

Apply the first input of YselY_{\mathrm{sel}} and set YprevY_{\mathrm{prev}} to its remaining sequence

Store {z(b)}b=1B\{z^{(b)}\}_{b=1}^{B}, βt+1\beta_{t+1}, YprevY_{\mathrm{prev}}, and nreusen_{\mathrm{reuse}}

desired Cartesian velocity supplied to a low-level controller. The diffusion model is trained on 96 human demonstrations: four for each of 24 routes, to generate joint state-input sequences and remains frozen during evaluation. It generates YY with N=8N=8 stages, conditioned on ct=stc_t=s_t. Only the first planned velocity u^0∣t\hat{u}_{0\mid t} is sent to the low-level controller, with replanning at every step. Since the demonstrations cover 24 routes, Pt\mathbb{P}_t is multimodal, and tilting suppresses routes blocked by the new obstacles. Deployment introduces additional forbidden regions absent during training. We use three layouts: top-left (one halfspace and one circle), top-right (its reflection about x=0.5x=0.5), and both (two halfspaces and one circle). We compare TCA-DPC and DPCC without constraint tightening (γ=0\gamma=0) and with a buffer γ>0\gamma>0, which shifts halfspace boundaries and enlarges circular obstacles. Figure 1 illustrates the demonstrations and constraint layouts. The constraint residuals gℓ,t(Y)≤0g_{\ell,t}(Y)\le0

Avoiding environment

Figure 1. Avoiding environment. (a) The 96 training demonstrations, coloured by route, i.e., by the side on which they pass the first obstacle and the gap they take between the two middle obstacles. (b)–(d) The three test-time constraint layouts: forbidden regions (blue) and their tightening by γ\gamma (grey). None of the forbidden regions appears in the training data.

specify obstacle avoidance, input bounds derived from the demonstrations, and consistency with an approximate dynamics model. Their explicit expressions, the sequence representation, and the implementation of the penalty Φt\Phi_t are provided in the supplementary material. We implement Algorithm 1 with pretrained diffusion models, selecting the verified sequence closest to the unexecuted portion of the previous plan. Further, we set the consecutive reuse limit to Rmax⁡=3R_{\max}=3 and the additional refinement budget to Mextra=10M_{\mathrm{extra}}=10. These limits were selected by manual tuning in preliminary simulations, considering closed-loop performance and computational cost. Model and training details, sampling settings, verification tolerances, and feedback parameters are provided in the supplementary material.

Baselines and metrics. All methods use the same pretrained diffusion models. Prior samples without verification, as in Diffuser [11]. Prior + verification is Alg. 1 with M=Mextra=0M=M_{\mathrm{extra}}=0 and the same warm start (one fresh z(b)z^{(b)} per step), verification, selection, and YprevY_{\mathrm{prev}} reuse; its candidates are thus step-wise correlated samples of Pt\mathcal{P}_t, and it differs from TCA-DPC only in omitting MALA refinement. Fixed β\beta uses β=β0\beta=\beta_0 (κ=0\kappa=0). DPCC-C and DPCC-T select sequences using cumulative projection cost and temporal consistency, respectively (see Römer et al. [19] for details). The maximum episode length is 200 control steps. Within this limit, we report goal success (G), defined as the fraction of episodes in which the end-effector reaches the goal line, and goal-and-constraint success (G&C), the fraction that reach the goal line without violating the original state constraints at any control step. Constraint satisfaction is evaluated without the tightening buffer, irrespective of the value of γ\gamma used by the controller. We also report the mean number of violating control steps per episode (Viol.) and the mean number of steps to the goal over goal-reaching episodes (Steps). For TCA-DPC and its baselines, FB is the fraction of control steps that reach the fallback branch of Alg. 1 (no verified sequence after the MextraM_{\mathrm{extra}} additional updates), and fˉ\bar{f} is the mean of ftf_t; both are pooled over each model’s control steps and averaged over models. Each layout is evaluated using five trained models and ten episodes per model, with standard errors computed across models. Each run uses an NVIDIA V100, A100 or H100 GPU.

Results

We evaluate the adapted sampler in closed loop on the three constraint layouts, comparing success and constraint violations against the untilted (pretrained diffusion prior), fixed-tilt (constant β\beta), and DPCC baselines. We then examine how the feedback rule adjusts the tilt strength during an episode, and how performance changes with constraint tightening γ\gamma. Table 1 compares the controllers without constraint tightening. Averaged over the three layouts, TCA-DPC achieves 82.7% G&C, compared with 61.3% for the untilted verification baseline and 71.3% for DPCC-C. The mean number of violating steps per episode decreases from 3.17 for the untilted verification baseline to 0.73 for our method. The largest difference from DPCC-C occurs in the both layout, where TCA-DPC achieves 74% success compared with 46%. These results are obtained without projection or constraint tightening (γ=0\gamma=0). Across the three layouts, TCA-DPC achieves the highest or joint-highest mean G&C among the evaluated controllers. Figure 2(a) compares the closed-loop end-effector paths obtained with the pretrained sampler and TCA-DPC in the top-left layout, where G&C increases from 10% to 94%.

Table 1: Closed-loop performance on the three constraint layouts without tightening (γ=0\gamma= 0); Viol., Steps and FB are averaged over layouts. Bold marks the best result and underlining the best among methods without projection.

MethodProj.Top-left GTop-left G&CTop-right GTop-right G&CBoth GBoth G&CViol. ↓Steps ↓FB
Prior✗1.000.10±.031.000.06±.021.000.02±.0216.6674.8–
Prior + verification✗1.000.90±.000.800.46±.060.940.48±.073.1773.10.12
Fixed β=β0\beta= \beta_0✗1.000.94±.040.600.38±.090.960.62±.112.1269.10.08
TCA-DPC✗1.000.94±.040.860.80±.050.940.74±.070.73<u>68.8</u>0.05
DPCC-C✓0.980.90±.030.940.78±.070.900.46±.110.8966.9–
DPCC-T✓0.940.68±.040.940.68±.070.860.56±.111.8565.7–

Table 1.

Figure 2. Top-left layout. (a) Closed-loop paths of the prior (grey) and TCA-DPC (blue) without tightening. (b) Tilt strength βt\beta_t after the first verification failure (median and IQR). (c) Goal-and-constraint success versus tightening γ\gamma.

Figure 2(b) illustrates online adaptation in the top-left layout without tightening. Following the first verification failure, the median tilt strength rises to the order of 10310^3 and subsequently decreases. These values emerge from verification feedback during execution. Selecting a suitable fixed strength beforehand may instead require separate calibration. Holding the tilt fixed at β=β0=10\beta=\beta_0=10 instead lowers the average goal-and-constraint success to 64.7% (Table 1) across layouts. Additional fixedtilt comparisons, sensitivity results, and adaptation plots for the other layouts are provided in the supplementary material. Figure 2(c) shows the effect of tightening in the top-left layout. Even without tightening (γ=0\gamma= 0), TCA-DPC achieves 94% goal-and-constraint success, increasing to 100% at γ=0.5\gamma= 0.5 cm and all larger displayed values. Without tightening, it also achieves 80% and 74% success in the top-right and both layouts, respectively, with higher mean G&C than both DPCC variants in each case. These additional results are reported in the supplementary material. Recorded mean computation times per control step are 1.0 s for TCA-DPC and approximately 0.7 s for DPCC. The additional cost arises from repeated denoising, constraint evaluation, and gradient computations within MALA. TCA-DPC nevertheless avoids projection and can operate without constraint tightening (γ=0\gamma= 0). Reducing this computational cost is a direction for future work.

Conclusion

We proposed a projection-free method that adapts a pretrained diffusion prior to test-time safety constraints through exponential tilting, MALA, and verification feedback. D3IL experiments showed improved goal-and-constraint success over untilted baselines, including without constraint tightening.

References

Proof for Proposition 1

Since Φt(Y)=0\Phi_t(Y)=0 on AtA_t, for every measurable A⊆YA\subseteq\mathcal{Y},

Qβ,t(A∣At)=Qβ,t(A∩At)Qβ,t(At)=Pt(A∩At)/Zβ,tPt(At)/Zβ,t=Pt(A∣At).Q_{\beta,t}(A\mid A_t)=\frac{Q_{\beta,t}(A\cap A_t)}{Q_{\beta,t}(A_t)}=\frac{P_t(A\cap A_t)/Z_{\beta,t}}{P_t(A_t)/Z_{\beta,t}}=P_t(A\mid A_t).

This proves preservation of the conditional distribution. Furthermore, Zβ,t=Pt(At)+∫Atce−βΦt(Y)Pt(dY)Z_{\beta,t}=P_t(A_t)+\int_{A_t^c}e^{-\beta\Phi_t(Y)}P_t(dY). The integral is non-increasing in β\beta and converges to zero by dominated convergence, since Φt>0\Phi_t>0 on AtcA_t^c. Thus, Zβ,t↓Pt(At)>0Z_{\beta,t}\downarrow P_t(A_t)>0, which gives Qβ,t(Atc)=1−Pt(At)Zβ,t↓0Q_{\beta,t}(A_t^c)=1-\frac{P_t(A_t)}{Z_{\beta,t}}\downarrow0 as β→∞\beta\to\infty.

Implementation details

Sequences, constraint residuals, and penalty. Each of the N=8N=8 stages of YY, the sequence generated by the diffusion model, contains the predicted current and desired end-effector positions, s^k∣t=(pk∣t,pk∣tdes)\hat{s}_{k|t}=(p_{k|t},p^{\mathrm{des}}_{k|t}), and the input u^k∣t\hat{u}_{k|t}. The residuals gℓ,tg_{\ell,t} comprise, for k=1,…,N−1k=1,\ldots,N-1 (the initial stage is fixed to s^0∣t=st\hat{s}_{0|t}=s_t and not checked), ajTpk∣t−bj+γa_j^\mathsf{T}p_{k|t}-b_j+\gamma, (r+γ)2−∥pk∣t−c∥2(r+\gamma)^2-\lVert p_{k|t}-c\rVert^2, for the halfspaces (∥aj∥=1\lVert a_j\rVert=1) and the circle, together with the input bounds for k=0,…,N−1k=0,\ldots,N-1, given by the smallest box that contains all demonstrated inputs, and the consistency with an explicit Euler model in which both the current and desired positions integrate the commanded velocity, pk+1∣t=pk∣t+Δt u^k∣tp_{k+1|t}=p_{k|t}+\Delta t\,\hat{u}_{k|t} and pk+1∣tdes=pk∣tdes+Δt u^k∣tp^{\mathrm{des}}_{k+1|t}=p^{\mathrm{des}}_{k|t}+\Delta t\,\hat{u}_{k|t} with Δt=1\Delta t=1 control step, for k=0,…,N−2k=0,\ldots,N-2, with residual rk∣tdynr^{\mathrm{dyn}}_{k|t} being the difference of the two sides. All residuals are in physical units (metres). The penalty is Φt(Y)=∑ℓρ(gℓ,t(Y)+m)\Phi_t(Y)=\sum_\ell\rho(g_{\ell,t}(Y)+m) with margin m=10−3m=10^{-3} and ρ(a)=0\rho(a)=0 for a≤0a\leq0, a2/(2η)a^2/(2\eta) for 0<a≤η0<a\leq\eta, and a−η/2a-\eta/2 for a>ηa>\eta, where η=10−3\eta=10^{-3} (ρ\rho is C1C^1, where C1C^1 is set of all continuously differentiable function).

Diffusion prior and adapted sampler. The prior is the pretrained diffusion model: a DDPM with a 1D temporal U-Net (Janner et al., [11]), K=20K=20 denoising steps, and sequences of N=8N=8 stages containing states and inputs. We train five models with different seeds on the demonstrations for 10510^5 steps with batch size 32 and learning rate 10−410^{-4}, use the checkpoint with the lowest test loss, and keep the models fixed during deployment. At each control step we generate B=4B=4 sequences. One of the BB noise vectors is drawn afresh and the others are reused from the previous control step. Each is refined with M=5M=5 MALA steps of size h=0.05h=0.05. A sequence passes verification when its largest state-constraint residual is at most ϵ=10−5\epsilon=10^{-5}, its largest dynamics residual satisfies ∥rk∣tdyn∥∞≤ϵdyn=0.025 m\lVert r^{\mathrm{dyn}}_{k|t}\rVert_\infty\leq\epsilon_{\mathrm{dyn}}=0.025\,\mathrm{m}, and its largest input-bound residual is at most ϵu=\epsilon_u= 3⋅10−33 \cdot10^{-3}. Stage k=0k = 0 is excluded from the state check because s^0∣t=st\hat{s}_{0|t} = s_t is imposed rather than generated, so it cannot be changed by refinement. The state constraints encode safety and are therefore verified strictly; a looser input tolerance admits the small overshoots of the planned velocities that the prior produces near the goal. The verification tolerances and the margin used in the penalty are implementation choices applied to the constraint residuals. They are distinct from the geometric tightening parameter γ\gamma, which shifts halfspace boundaries and enlarges circular obstacles. In the experiments, the failure fraction is computed using these verification tolerances. Proposition 1 concerns the exact zero-penalty set of the theoretical formulation; the implemented acceptance test need not coincide with this set. Among the verified sequences, we select the minimizer of Jsel(Y)=∥(s^k∣t−s^k+1prev)k=0N−2∥2J_{\mathrm{sel}}(Y) = \left\|(\hat{s}_{k|t} - \hat{s}^{\mathrm{prev}}_{k+1})_{k=0}^{N-2}\right\|_2 over normalized states, where YprevY_{\mathrm{prev}} always has NN stages (a reused sequence is shifted, repeating its last stage); at the first step, the first verified sequence is taken. The tilt strength starts at β0=10\beta_0 = 10 and is updated by (13) with κ=500\kappa= 500, δ=0.25\delta= 0.25, and βmax⁡=104\beta_{\max} = 10^4.

Additional results and ablation study

Additional results

MethodProj.Top-left GTop-left G&CTop-right GTop-right G&CBoth GBoth G&CViol. ↓Steps ↓FB
Prior×\times1.000.10±.031.000.06±.021.000.02±.0216.6674.8–
Prior + verification×\times0.980.50±.090.700.14±.040.940.38±.116.7574.40.31
TCA-DPC×\times1.001.00±.000.98<u>0.98±.02</u>0.94<u>0.90±.04</u><u>0.07</u><u>68.1</u>0.13
DPCC-C✓\checkmark1.001.00±.001.001.00±.001.001.00±.000.0063.8–
DPCC-T✓\checkmark1.001.00±.001.000.96±.020.940.94±.060.0464.3–

Table 2. Closed-loop results at γ=2.5\gamma= 2.5 cm, mean (± s.e.) over 5 models ×\times 10 episodes. Viol., Steps and FB are averaged over layouts. Proj.: run-time projection. Bold: best; <u>underlined</u>: best without projection.

Table 2 reports results with constraint tightening γ=2.5\gamma= 2.5 cm. TCA-DPC achieves 100% and 98% goal-and-constraint success in the top-left and top-right layouts, respectively, comparable to the projection-based methods. The gap is larger in the more restrictive both layout, where TCA-DPC achieves 90%, compared with 100% for DPCC-C and 94% for DPCC-T. Across all layouts, TCA-DPC substantially improves upon the untilted verification baseline. The relative performance depends on constraint tightening. Without tightening (γ=0\gamma= 0), TCA-DPC achieves higher mean goal-and-constraint success than both DPCC variants in every layout (Table 1). Introducing γ=2.5\gamma= 2.5 cm improves both methods, with a larger gain for DPCC-C. This is consistent with the role of tightening in compensating for prediction errors: projection ensures nominal feasibility, but trajectories close to constraint boundaries remain vulnerable to differences between predicted and executed states. TCA-DPC therefore demonstrates effective adaptation without a buffer, while DPCC-C achieves higher success when the tested buffer is included.

Figure 3 extends the comparison in the main paper to all three test-time constraint layouts. Column (a) shows closed-loop trajectories without tightening (γ=0\gamma= 0): TCA-DPC trajectories largely avoid both the original obstacles and the additional forbidden regions, whereas trajectories generated by the prior frequently enter the latter. Column (b) shows that, in every layout, the tilt strength rises to the order of 10210^{2}–10310^{3} after the first verification failure, well below βmax⁡\beta_{\max}, and decreases once failures become rare, most clearly in the both layout. Column (c) shows that, without tightening (γ=0\gamma= 0), TCA-DPC achieves higher mean goal-and-constraint success than both DPCC variants in every layout. The largest difference occurs in the both layout, with 74% success for TCA-DPC, compared with 46% for DPCC-C and 56% for DPCC-T.

Figure 3. Results on the three constraint layouts, one per row. (a) Closed-loop paths of the prior (grey) and TCA-DPC (blue) without tightening. (b) Tilt strength βt\beta_t after the first verification failure, median and interquartile range. (c) Success rate versus tightening γ\gamma. Results over 50 episodes per layout.

Ablation study

On the both layout at γ=2.5\gamma= 2.5 cm, each row of Table 3 changes one parameter from the reference setting (Section B), except the fixed-tilt row, which uses β=103\beta= 10^3, and the M=0M = 0 row, which also sets Mextra=0M_{\mathrm{extra}} = 0. Most of the gain comes from the first few MALA steps, with little further improvement beyond M=5M = 5. In this experiment, adaptive tilting attains higher mean success than a fixed β=103\beta= 10^3 (κ=0\kappa= 0), and, for κ∈{100,500,2000}\kappa\in\{100, 500, 2000\}, mean success increases with κ\kappa. Success is robust to δ∈{0.05,0.25,0.5}\delta\in\{0.05, 0.25, 0.5\}, with fˉ≈0.25\bar{f} \approx0.25 in all cases. Raising βmax⁡\beta_{\max} did not change the reported performance.

ParameterValueG&C↑\uparrowViol.↓\downarrowfˉ\bar{f}FB
Reference (TCA-DPC)0.90±\pm.040.100.250.17
MALA steps MM (ref. 5)0 (prior + verification)0.38±\pm.111.840.280.24
10.78±\pm.080.260.240.18
20.76±\pm.050.500.260.19
100.88±\pm.060.120.260.18
Gain κ\kappa (ref. 500)0 (fixed β=103\beta= 10^{3})0.84±\pm.080.320.260.19
1000.76±\pm.070.620.250.20
20000.94±\pm.020.000.250.16
Budget δ\delta (ref. 0.25)0.050.96±\pm.020.060.240.17
0.50.94±\pm.040.140.250.18
Cap βmax⁡\beta_{\max} (ref. 10410^{4})10510^{5}0.90±\pm.040.100.250.17

Table 3. Ablations on the both layout at γ=2.5\gamma= 2.5 cm. fˉ\bar{f}: mean failure fraction.

References

  1. [1]Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Roni Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In International Conference on Learning Representations, volume 2024, pages 51304–51323, 2024.arxiv.org/abs/2302.07121
  2. [2]P. G. Bissiri, C. C. Holmes, and S.G. Walker. A general framework for updating belief distributions. Journal of the Royal Statistical Society Series B: Statistical Methodology, 78(5):1103–1130, 2016.
  3. [3]Sabrina Bodmer, Rahel Rickenbach, René Zurbrügg, Melanie N Zeilinger, and Colin Jones. MOSAIC: Mode-and support-aware inference-time control of robot diffusion models. arXiv preprint arXiv:2610.17001, 2026.arxiv.org/abs/2610.17001
  4. [4]Vladimir Igorevich Bogachev. Measure theory, volume 1. Springer, 2007.
  5. [5]Francesco Borrelli, Alberto Bemporad, and Manfred Morari. Predictive Control for Linear and Hybrid Systems. Cambridge University Press, 2017.DOI
  6. [6]João Carvalho, An T. Le, Mark Baierl, Dorothea Koert, and Jan Peters. Motion planning diffusion: Learning and planning of robot motions with diffusion models. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1916–1923, 2023.
  7. [7]Jacob K. Christopher, Stephen Baek, and Ferdinando Fioretto. Constrained synthesis with projected diffusion models. In Advances in Neural Information Processing Systems, volume 37, pages 89307–89333, 2024.arxiv.org/abs/2402.03559
  8. [8]Yingqing Guo, Hui Yuan, Yukang Yang, Minshuo Chen, and Mengdi Wang. Gradient guidance for diffusion models: An optimization perspective. In Advances in Neural Information Processing Systems, volume 37, pages 90736–90770, 2024.arxiv.org/abs/2404.14743
  9. [9]Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, volume 33, pages 6840–6851, 2020.arxiv.org/abs/2006.11239
  10. [10]Tzu-Yuan Huang, Armin Lederer, Nicolas Hoischen, Jan Brudigam, Xuehua Xiao, Stefan Sosnowski, and Sandra Hirche. Toward near-globally optimal nonlinear model predictive control via diffusion models. In Proceedings of the 7th Annual Learning for Dynamics & Control Conference, volume 283 of Proceedings of Machine Learning Research, pages 777–790. PMLR, 2025.
  11. [11]Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 9902–9915. PMLR, 2022.arxiv.org/abs/2205.09991
  12. [12]Xiaogang Jia, Denis Blessing, Xinkai Jiang, Moritz Reuss, Atalay Donat, Rudolf Lioutikov, and Gerhard Neumann. Towards diverse behaviors: A benchmark for imitation learning with human demonstrations. In International Conference on Learning Representations, volume 2024, pages 41666–41686, 2024.
  13. [13]Pau Marquez Julbe, Julian Nubert, Henrik Hose, Sebastian Trimpe, and Katherine J Kuchenbecker. Diffusion-based approximate MPC: Fast and consistent imitation of multi-modal action distributions. In 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5633–5640. IEEE, 2025.arxiv.org/abs/2504.04603
  14. [14]Tanmay Khandait, Preetom Biswas, Hideki Okamoto, Bardh Hoxha, Georgios Fainekos, and Giulia Pedrielli. Diffusion-guided search via exponential tilting (DiffTilt): An application to falsification of safety-critical systems. arXiv preprint arXiv:2607.23134, 2026.arxiv.org/abs/2607.23134
  15. [15]Yin Tat Lee, Aaron Sidford, and Santosh S. Vempala. Efficient convex optimization with membership oracles. In Proceedings of the 31st Conference On Learning Theory, volume 75 of Proceedings of Machine Learning Research, pages 1292–1294, 2018.arxiv.org/abs/1706.07357
  16. [16]Zakaria Mhammedi. Efficient projection-free online convex optimization with membership oracle. In Proceedings of Thirty Fifth Conference on Learning Theory, volume 178 of Proceedings of Machine Learning Research, pages 5314–5390, 2022.arxiv.org/abs/2111.05818
  17. [17]Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, and Jan Peters. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 7(1-2): 1–179, 2018.
  18. [18]G.O. Roberts and O. Stramer. Langevin diffusions and Metropolis-Hastings algorithms. Methodology and computing in applied probability, 4(4):337–357, 2002.DOI
  19. [19]Ralf Römer, Alexander von Rohr, and Angela Schoellig. Diffusion predictive control with constraints. In Proceedings of the 7th Annual Learning for Dynamics & Control Conference, volume 283 of Proceedings of Machine Learning Research, pages 791–803. PMLR, 2025.arxiv.org/abs/2412.09342
  20. [20]Yibo Shi, Braghadeesh Lakshminarayanan, and Cristian R Rojas. A Metropolis-adjusted Langevin algorithm for sampling Jeffreys prior. In 2025 IEEE 64th Conference on Decision and Control (CDC), pages 2359–2364, 2025.arxiv.org/abs/2504.06372
  21. [21]Sreevardhan Sirigiri, Weiming Zhi, and Fabio Ramos. BayesFP: Posterior estimation for flow-based policies via Feynman-Kac sampling. arXiv preprint arXiv:2606.21014, 2026.arxiv.org/abs/2606.21014
  22. [22]Marta Skreta, Tara Akhound-Sadegh, Viktor Ohanesian, Roberto Bondesan, Alan Aspuru-Guzik, Arnaud Doucet, Rob Brekelmans, Alexander Tong, and Kirill Neklyudov. Feynman-Kac correctors in diffusion: Annealing, guidance, and product of experts. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 55906–55949. PMLR, 2025.arxiv.org/abs/2503.02819
  23. [23]Julen Urain, Ajay Mandlekar, Yilun Du, Mahi Shafullah, Danfei Xu, Katerina Fragkiadaki, Georgia Chalvatzaki, and Jan Peters. Deep generative models in robotics: A survey on learning from multimodal demonstrations. arXiv preprint arXiv:2408.04380, 2024.arxiv.org/abs/2408.04380

Paper details

vlnc:W_y7k5xtpy
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents