DiffPlanner: Adaptive Online Diffusion Planning

• DiffPlanner is an adaptive online replanning framework that utilizes diffusion models and likelihood-based criteria to decide when to repair or replace trajectories.
• It features two replanning operators—‘replan with future context’ and ‘replan from scratch’—to balance repair efficiency with trajectory fidelity.
• Empirical results on Maze2D, stochastic locomotion, and robotic control benchmarks demonstrate significant performance gains over fixed-interval replanning.

Searching arXiv for the primary paper and related usages of “DiffPlanner” to ground the article in current literature. DiffPlanner most commonly denotes the adaptive online replanning framework introduced in “Adaptive Online Replanning with Diffusion Models,” where a diffusion planner is converted from a one-shot trajectory generator into an online control layer that monitors whether the currently executed plan still appears feasible, decides when replanning should occur, and edits existing trajectories rather than discarding them outright (Zhou et al., 2023). In this formulation, the central question is not only how to plan with diffusion models, but when to replan, and how to preserve useful structure—especially the original goal state—during repair. The name is also context-dependent in later literature: it reappears in autonomous driving for a transformer-based diffusion planner with flexible guidance (Zheng et al., 26 Jan 2025) and for an open-source modular benchmark inside Autoware (Li et al., 1 Mar 2026).

1. Problem setting and motivation

DiffPlanner is posed in a standard trajectory-optimization setting over state-action trajectories

$\tau=(s_0,a_0,s_1,a_1,\ldots,s_T,a_T),$

with stochastic dynamics

$s_{t+1}=f_\sigma(s_t,a_t),$

and an objective of finding an action sequence maximizing expected return,

$$\arg\max_{a_{0:T}} \mathcal{J}(s_0,a_{0:T}) = \arg\max_{a_{0:T}} \mathbb{E}_{s_{1:T}}\Big[\sum_{t=0}^T r(s_t,a_t)\Big].$$

As in Diffuser and Decision Diffuser, the planner operates over state-action trajectories and actions are recovered via inverse dynamics from the planned states (Zhou et al., 2023).

The motivation for adaptive replanning is specific to diffusion planners. They are strong at generating complete trajectories, but those trajectories often become invalid during execution because of stochastic transitions, action noise, partial observability, or external changes. Direct open-loop execution causes small errors to accumulate. Replanning naively at every timestep is also undesirable because diffusion sampling is expensive, and frequent switching between newly sampled plans can prevent consistent progress toward any particular goal. DiffPlanner addresses this tension by treating replanning as an online decision problem rather than a fixed schedule.

This framing shifts diffusion planning from single-shot sequence generation to closed-loop control. The diffusion model is used not only to synthesize candidate trajectories, but also to assess whether the current plan still lies on a plausible trajectory manifold for the task.

2. Likelihood-based replanning criterion

The central mechanism for deciding when to replan is a likelihood estimate of the partially executed trajectory under the diffusion model. Let $\tau_{\to k}$ denote the trajectory after replacing executed planned states with observed states: $$\tau_{\to k} = (s_0,a_0,\bar{s}_1,\bar{a}_1,\dots,\bar{s}_k,\bar{a}_k,s_{k+1},\dots,s_T,a_T).$$ The underlying intuition is that a feasible trajectory should be easy for the diffusion model to denoise or explain, whereas an infeasible one should have low model likelihood (Zhou et al., 2023).

The score is constructed using the same variational denoising objective as Diffuser. For diffusion step $i$, the method corrupts the trajectory with the forward process $q(\tau^i \mid \tau_{\to k})$, then compares the posterior denoising distribution with the model’s reverse transition using

$$D_{\mathrm{KL}}\!\left(q(\tau^{i-1}\mid \tau^i,\tau_{\to k}) \;\|\; p_\theta(\tau^{i-1}\mid \tau^i)\right).$$

This is evaluated at several fixed diffusion depths $i\in I$ and averaged into

$$L_t = \frac{1}{|I|}\sum_{i\in I} D_{\mathrm{KL}}\!\left(q(\tau^{i-1}\mid \tau^i,\tau^0) \;\|\; p_\theta(\tau^{i-1}\mid \tau^i)\right).$$

The decision policy is a two-threshold scheme.

Outcome	Trigger	Action
Continue	Otherwise	Keep executing current plan
Replan with future	$s_{t+1}=f_\sigma(s_t,a_t),$0	Cheap local repair
Replan from scratch	$s_{t+1}=f_\sigma(s_t,a_t),$1 or fixed horizon interval reached	Full reverse-diffusion resampling

In the algorithm, there are two thresholds: a stricter one $s_{t+1}=f_\sigma(s_t,a_t),$2 for “replan from scratch” and a looser one $s_{t+1}=f_\sigma(s_t,a_t),$3 for “replan with future” (with $s_{t+1}=f_\sigma(s_t,a_t),$4). The thresholds used in experiments are $s_{t+1}=f_\sigma(s_t,a_t),$5 for replan-from-scratch and $s_{t+1}=f_\sigma(s_t,a_t),$6 for replan-with-future. Likelihood is estimated using diffusion steps $s_{t+1}=f_\sigma(s_t,a_t),$7 for Maze2D and stochastic environments, and $s_{t+1}=f_\sigma(s_t,a_t),$8 for RLBench (Zhou et al., 2023).

This criterion is the main difference from fixed-interval replanning. Rather than assuming that deviation grows uniformly with time, DiffPlanner asks whether the current plan remains explainable under the learned generative model.

3. Replanning operators and online execution

DiffPlanner provides two primary replanning operators. The first is “replan from scratch.” It ignores the old sample, initializes

$s_{t+1}=f_\sigma(s_t,a_t),$9

and runs the full reverse diffusion process for $$\arg\max_{a_{0:T}} \mathcal{J}(s_0,a_{0:T}) = \arg\max_{a_{0:T}} \mathbb{E}_{s_{1:T}}\Big[\sum_{t=0}^T r(s_t,a_t)\Big].$$0 steps,

$$\arg\max_{a_{0:T}} \mathcal{J}(s_0,a_{0:T}) = \arg\max_{a_{0:T}} \mathbb{E}_{s_{1:T}}\Big[\sum_{t=0}^T r(s_t,a_t)\Big].$$1

while enforcing the task conditions $$\arg\max_{a_{0:T}} \mathcal{J}(s_0,a_{0:T}) = \arg\max_{a_{0:T}} \mathbb{E}_{s_{1:T}}\Big[\sum_{t=0}^T r(s_t,a_t)\Big].$$2 at each denoising step. This is the most expressive option, but also the most expensive (Zhou et al., 2023).

The second, and preferred, mechanism is “replan with future context.” Here the method repairs only the future portion of the current trajectory while preserving the already executed prefix and the original goal information carried by the tail. If $$\arg\max_{a_{0:T}} \mathcal{J}(s_0,a_{0:T}) = \arg\max_{a_{0:T}} \mathbb{E}_{s_{1:T}}\Big[\sum_{t=0}^T r(s_t,a_t)\Big].$$3 denotes the partially executed plan, the method constructs

$$\arg\max_{a_{0:T}} \mathcal{J}(s_0,a_{0:T}) = \arg\max_{a_{0:T}} \mathbb{E}_{s_{1:T}}\Big[\sum_{t=0}^T r(s_t,a_t)\Big].$$4

padding the remainder with repeated copies of the last predicted state-action so that the input length matches the diffusion model. It then applies the forward noising process for only $$\arg\max_{a_{0:T}} \mathcal{J}(s_0,a_{0:T}) = \arg\max_{a_{0:T}} \mathbb{E}_{s_{1:T}}\Big[\sum_{t=0}^T r(s_t,a_t)\Big].$$5 steps,

$$\arg\max_{a_{0:T}} \mathcal{J}(s_0,a_{0:T}) = \arg\max_{a_{0:T}} \mathbb{E}_{s_{1:T}}\Big[\sum_{t=0}^T r(s_t,a_t)\Big].$$6

followed by reverse denoising for $$\arg\max_{a_{0:T}} \mathcal{J}(s_0,a_{0:T}) = \arg\max_{a_{0:T}} \mathbb{E}_{s_{1:T}}\Big[\sum_{t=0}^T r(s_t,a_t)\Big].$$7 steps with the same conditioning $$\arg\max_{a_{0:T}} \mathcal{J}(s_0,a_{0:T}) = \arg\max_{a_{0:T}} \mathbb{E}_{s_{1:T}}\Big[\sum_{t=0}^T r(s_t,a_t)\Big].$$8: $$\arg\max_{a_{0:T}} \mathcal{J}(s_0,a_{0:T}) = \arg\max_{a_{0:T}} \mathbb{E}_{s_{1:T}}\Big[\sum_{t=0}^T r(s_t,a_t)\Big].$$9

This operator functions as trajectory editing. It keeps the executed prefix fixed through observation replacement, rewrites the future, and retains goal information encoded in the tail. The paper also discusses “replan from previous context,” which keeps the executed prefix and locally repairs the rest, but empirically finds future-context replanning to work better (Zhou et al., 2023).

At test time, the complete loop is: sample a fresh trajectory; execute one step; replace the planned next state with the observed state; compute the likelihood score $\tau_{\to k}$0; then either keep executing, locally repair the tail, or fully resample depending on the score and schedule. The resulting planner is therefore an online controller with a built-in feasibility detector.

4. Empirical behavior and implementation profile

The strongest empirical gains appear on long-horizon planning benchmarks. On Maze2D, the reported scores for the adaptive replanning method are $\tau_{\to k}$1 on U-Maze, $\tau_{\to k}$2 on Medium, and $\tau_{\to k}$3 on Large, compared with Diffuser’s $\tau_{\to k}$4, $\tau_{\to k}$5, and $\tau_{\to k}$6, respectively. The paper summarizes this as a $\tau_{\to k}$7 gain over Diffuser on average across Maze2D tasks, and about $\tau_{\to k}$8 on the Large setting (Zhou et al., 2023). Qualitatively, the method replans around walls and infeasible branches, whereas fixed-interval replanning can continue following a bad trajectory and waste steps.

On stochastic control tasks, the method is evaluated on a stochastic version of D4RL locomotion with injected transition randomness. Across Hopper and Walker2d variants, it outperforms Decision Diffuser and is competitive with or better than strong offline RL baselines; the reported average normalized return is $\tau_{\to k}$9 for the method versus $$\tau_{\to k} = (s_0,a_0,\bar{s}_1,\bar{a}_1,\dots,\bar{s}_k,\bar{a}_k,s_{k+1},\dots,s_T,a_T).$$0 for Decision Diffuser (Zhou et al., 2023). On RLBench robotic control, across open box, close box, and close fridge, the average success rate is $$\tau_{\to k} = (s_0,a_0,\bar{s}_1,\bar{a}_1,\dots,\bar{s}_k,\bar{a}_k,s_{k+1},\dots,s_T,a_T).$$1, above Decision Diffuser’s $$\tau_{\to k} = (s_0,a_0,\bar{s}_1,\bar{a}_1,\dots,\bar{s}_k,\bar{a}_k,s_{k+1},\dots,s_T,a_T).$$2 and far above BC, IQL, and DT baselines.

Ablations are integral to the method’s interpretation. Fixed-interval replanning can help initially but hurts when it becomes too frequent, because the agent never commits to a coherent long-horizon plan. The likelihood-triggered policy also outperforms simpler thresholding based on state distance or past context only. Among the tested repair strategies, “replan with future context” performs best (Zhou et al., 2023).

Implementation follows the Diffuser or Decision Diffuser temporal U-Net: six residual blocks, temporal convolutions, GroupNorm, and Mish activations. Conditioning and timestep embeddings are 128-dimensional MLP outputs. For RLBench, the visual pipeline uses a pretrained CLIP image encoder followed by a 2-layer MLP. Training uses Adam with learning rate $$\tau_{\to k} = (s_0,a_0,\bar{s}_1,\bar{a}_1,\dots,\bar{s}_k,\bar{a}_k,s_{k+1},\dots,s_T,a_T).$$3, batch size $$\tau_{\to k} = (s_0,a_0,\bar{s}_1,\bar{a}_1,\dots,\bar{s}_k,\bar{a}_k,s_{k+1},\dots,s_T,a_T).$$4, and $$\tau_{\to k} = (s_0,a_0,\bar{s}_1,\bar{a}_1,\dots,\bar{s}_k,\bar{a}_k,s_{k+1},\dots,s_T,a_T).$$5 steps. Planning horizon is $$\tau_{\to k} = (s_0,a_0,\bar{s}_1,\bar{a}_1,\dots,\bar{s}_k,\bar{a}_k,s_{k+1},\dots,s_T,a_T).$$6–$$\tau_{\to k} = (s_0,a_0,\bar{s}_1,\bar{a}_1,\dots,\bar{s}_k,\bar{a}_k,s_{k+1},\dots,s_T,a_T).$$7 in Maze2D or Multi2D and $$\tau_{\to k} = (s_0,a_0,\bar{s}_1,\bar{a}_1,\dots,\bar{s}_k,\bar{a}_k,s_{k+1},\dots,s_T,a_T).$$8 in stochastic locomotion and RLBench. Full replanning uses substantially more diffusion steps than repair—for example, $$\tau_{\to k} = (s_0,a_0,\bar{s}_1,\bar{a}_1,\dots,\bar{s}_k,\bar{a}_k,s_{k+1},\dots,s_T,a_T).$$9 versus $i$0 in Maze2D, $i$1 versus $i$2 in stochastic locomotion, and $i$3 versus $i$4 in RLBench (Zhou et al., 2023).

5. Assumptions, limitations, and common misunderstandings

A common misunderstanding is to treat DiffPlanner as a new base diffusion planner. The paper states the opposite: the overall contribution is not a new planner per se, but an adaptive online control layer over diffusion planners that decides when to trust the current plan and how to edit it efficiently when trust is lost (Zhou et al., 2023).

Its main assumptions are correspondingly explicit. First, the method depends heavily on the quality of the learned diffusion planner: if the model itself does not capture the task distribution well, its likelihood signal may be unreliable. Second, the likelihood-based criterion can miss some failures; the authors note that some planning failures are insensitive to this score, suggesting that it may be beneficial to combine it with other replanning signals. Third, while local repair is cheaper, it can fail when the current trajectory is too corrupted, in which case a full replan from scratch is necessary.

The principal methodological debate surrounding the framework concerns adaptivity versus regularity. Fixed-interval replanning offers simplicity, but the paper’s ablations show that excessive frequency destabilizes long-horizon commitment. DiffPlanner argues for a model-based notion of feasibility instead: the diffusion model itself becomes the detector for when a plan should no longer be trusted. This suggests a different interpretation of diffusion planning, in which denoising likelihood is not merely a training objective but also an online control signal.

6. Later developments and broader uses of the name

Subsequent work extends or reframes different components of the DiffPlanner agenda. AdaptDiffuser turns diffusion planning into a self-evolving process that generates synthetic expert data under reward guidance and finetunes the planner for better seen-task and unseen-task generalization (Liang et al., 2023). LoMAP focuses on reliability, arguing that guided diffusion planners fail because of manifold deviation and proposing a training-free local projection module that snaps guided samples back to a low-rank approximation of the offline trajectory manifold (Lee et al., 1 Jun 2025). Temporal Diffusion Planner distributes denoising effort across the time dimension, reuses noisy long-horizon plans, and adds an automated replanning mechanism to improve decision frequency by $i$5–$i$6 times while maintaining higher or comparable performance on D4RL (Guo et al., 26 Nov 2025).

Other lines of work modify the temporal structure of diffusion planning rather than the replanning trigger itself. Mixed Density Diffuser uses non-uniform temporal resolution, replacing a fixed step size with a sequence of tunable jump sizes $i$7 and reporting state-of-the-art results on Maze2D, Franka Kitchen, and Antmaze (Stambaugh et al., 27 Oct 2025). VH-Diffuser treats the horizon as an instance-specific variable predicted from the start-goal pair, thereby addressing length mismatch in goal-conditioned planning (Liu et al., 15 Sep 2025).

In autonomous driving, the name “DiffPlanner” is used in distinct but related senses. “Diffusion-Based Planning for Autonomous Driving with Flexible Guidance” denotes a transformer-based diffusion planner that jointly models prediction and planning and supports training-free classifier guidance for safety, comfort, speed, and drivable-area constraints (Zheng et al., 26 Jan 2025). “An Open-Source Modular Benchmark for Diffusion-Based Motion Planning in Closed-Loop Autonomous Driving” uses DiffPlanner to refer to a modular benchmark that decomposes a monolithic diffusion planner into three executable ONNX modules, reimplements the DPM-Solver++ loop in native C++, and demonstrates a $i$8 latency reduction from encoder caching in Autoware or AWSIM closed-loop evaluation (Li et al., 1 Mar 2026). This multiplicity of usage means that the term now denotes a family of diffusion-based planning ideas rather than one uniquely fixed architecture.

Taken together, these developments suggest that the original adaptive online replanning formulation occupies a specific position within a broader research program: it addresses the online trust problem of diffusion planners—when to retain, repair, or replace a plan—while later work addresses orthogonal issues such as manifold fidelity, temporal allocation, variable horizon selection, closed-loop deployment, and real-time systems integration.