Hierarchical Diffusion Planner

• Hierarchical Diffusion Planner is a planning framework that splits decision-making into high-level subgoal generation and low-level detailed execution for complex tasks.
• It leverages diverse designs such as diffusion-guided waypoint planning and coupling with reinforcement learning or flow-based policies to improve planning efficiency.
• Empirical results show significant gains in task success and speedup, with successful applications in motion planning, navigation, and video generation benchmarks.

Hierarchical diffusion planner denotes a class of planning architectures in which diffusion models are assigned distinct roles across levels of abstraction, typically separating coarse, long-horizon structure from short-horizon feasibility or execution. In the cited literature, this hierarchy appears as a high-level “jumpy” sub-goal generator paired with a low-level segment diffuser, a diffusion global planner paired with a reinforcement-learning controller, an obstacle-aware planner paired with a short-horizon relative-motion diffusion policy, or a planner–executor stack for autoregressive video diffusion (Chen et al., 2024, Wu et al., 2024, Ravan et al., 2024, Wang et al., 26 Jun 2026). The common design objective is not a single algorithmic template, but a recurrent decomposition: sparse waypoints, latent subgoals, options, key states, or step prompts are generated at one level, and dense trajectories, actions, or frame chunks are synthesized at another.

1. Core concept and scope

In “Simple Hierarchical Planning with Diffusion,” the Hierarchical Diffuser splits planning into two levels of diffusion models: a High-Level “Jumpy” Planner, or Sparse Diffuser, operating on a subsampled sequence of states, and a Low-Level Diffusion Planner that connects adjacent sub-goal pairs with short segments (Chen et al., 2024). A closely related formulation appears in long-horizon motion planning, where the upper level predicts sparse plans and the lower level fills in continuous trajectories, as in cascaded diffusion motion planning and multiscale diffusion planning (Sharma et al., 21 May 2025, Chen et al., 25 Mar 2025).

Other systems preserve the high-level/low-level split while changing the lower-level mechanism. In the pursuit–evasion planner, the top level is a Diffusion-based Global Planner over sparse 2D waypoints and the bottom level is a Soft Actor-Critic local evasive policy over continuous 2D velocity commands (Wu et al., 2024). In PoPi, the hierarchy consists of an obstacle-aware waypoint planner and a short-horizon relative-motion diffusion policy for unknown object dynamics (Ravan et al., 2024). In HDFlow, the upper level is a high-level diffusion planner over latent subgoals, whereas the lower level is a rectified-flow planner for dense latent trajectories (Gireesh et al., 6 May 2026).

The same structural idea extends beyond classical robotics. SkillDiffuser uses a high-level skill abstraction module that learns discrete skill representations from visual observations and language instructions, then conditions a low-level diffusion model over future states on those skill embeddings (Liang et al., 2023). TempAct applies hierarchy to autoregressive video diffusion: a planner policy samples a temporal decomposition of step prompts, and an AR diffusion executor rolls out chunked latent-frame generation under smoothed prompts (Wang et al., 26 Jun 2026). This breadth suggests that “hierarchical diffusion planner” is best understood as a family of generative planning decompositions rather than a single named architecture.

2. Recurrent architectural patterns

Pattern	Representative systems	Hierarchical split
Coarse-to-fine temporal planning	HD, HM-Diffuser, Cascaded Diffusion	sparse subgoals or subsampled path $\rightarrow$ dense trajectory
Planner/controller decomposition	Diffusion-RL PEG, PoPi, HDFlow	global plan or latent subgoals $\rightarrow$ control or flow execution
Search-guided diffusion	XDiffuser, DOPPLER	graph or option search $\rightarrow$ guided denoising
Coupled hierarchical diffusion	CHD, TempAct, HD-ExpIt	HL proposals updated using LL feedback or on-policy success

A first pattern is temporal subsampling. Hierarchical Diffuser trains the upper level on a subsampled trajectory $\tau^{\mathrm{SD}}=[s_0,s_K,\dots,s_{HK}]$, then trains the lower level on contiguous short segments between consecutive subgoals (Chen et al., 2024). HM-Diffuser generalizes this into a hierarchy of $L$ diffusion planners operating at different temporal resolutions, with level $\ell$ planning subgoals spaced $j_\ell$ time-steps apart and lower levels refining the gaps between higher-level subgoals (Chen et al., 25 Mar 2025). Cascaded diffusion motion planning uses an analogous two-level design in which Level 2 operates on a subsampled path and Level 1 refines it using coarse subgoals and linearly interpolated reference points (Sharma et al., 21 May 2025).

A second pattern is explicit planner/controller separation. In partially observable pursuit–evasion, the diffusion model produces global waypoints while the SAC policy trades off waypoint tracking against avoidance of pursuers through an entropy-regularized objective (Wu et al., 2024). PoPi similarly lets the planner operate under an approximate holonomic rigid-motion model while the low-level diffusion policy learns short-horizon robot–chair–ground interactions from demonstrations (Ravan et al., 2024). HDFlow preserves the hierarchy but replaces the lower diffusion model with rectified flow to accelerate dense trajectory synthesis (Gireesh et al., 6 May 2026).

A third pattern uses an external scaffold. XDiffuser computes a plan over a state-space graph before denoising and uses the resulting waypoint scaffold to guide a single diffusion pass (Hassidof et al., 16 May 2026). DOPPLER generates a chain of short-horizon options with diffusion, but evaluates and selects them online with an option-value critic under co-safe LTL progression (Feng et al., 2024).

A fourth pattern increases coupling between levels rather than keeping them modular. CHD models high-level subgoals and low-level trajectories jointly within a unified diffusion process, with a shared classifier that passes low-level feedback upstream so that subgoals self-correct while sampling proceeds (Hao et al., 12 May 2025). HD-ExpIt does not add an explicit proxy model; instead it repeatedly retrains a hierarchical diffusion policy on successful on-policy rollouts, so that the high-level planner is trained only on subgoal sequences that the low-level controller could actually execute (Grislain et al., 5 Mar 2026).

3. Mathematical structure

Most hierarchical diffusion planners inherit the standard discrete-time diffusion backbone. A representative formulation is

$$q(x_m\mid x_0)=\mathcal{N}\bigl(x_m;\sqrt{\bar\alpha_m}\,x_0,\,(1-\bar\alpha_m)I\bigr), \qquad p_\theta(x_{m-1}\mid x_m)=\mathcal{N}\bigl(x_{m-1};\mu_\theta(x_m,m),\sigma_m^2 I\bigr),$$

with training by MSE noise prediction or denoising score matching (Chen et al., 2024). Hierarchy enters through what $x_0$ represents: a sparse subgoal sequence, a full trajectory segment, a latent option, or a plan-conditioned chunk.

In explicit two-level motion planners, the factorization is often written directly over levels. Cascaded diffusion models use

$$p(\tau^1,\tau^2\mid o,q_0,q_g) = p(\tau^2\mid o,q_0,q_g)\cdot p(\tau^1\mid o,q_0,q_g,\tau^2),$$

where $\rightarrow$0 is the short global plan and $\rightarrow$1 is the full-resolution path (Sharma et al., 21 May 2025). CHD introduces a joint planning distribution over high-level subgoals $\rightarrow$2 and low-level segments $\rightarrow$3 conditioned on binary optimality indicators:

$\rightarrow$4

This formulation makes the control-as-inference term explicit and is central to classifier-guided coupling between the two levels (Hao et al., 12 May 2025).

Some hierarchical planners alter not only the factorization but also the noise model. The task-conditioned uncertainty-aware planner begins with a Gaussian-process motion prior $\rightarrow$5, conditions it on sparse, task-centric key states treated as noisy observations, and thereby instantiates a task-conditioned structured Gaussian prior $\rightarrow$6 before trajectory denoising (Kim et al., 30 Sep 2025). The lower-level diffusion process then uses biased, non-isotropic corruption with mean $\rightarrow$7 and covariance $\rightarrow$8, and the denoiser is trained with a Mahalanobis DDPM loss weighted by $\rightarrow$9 (Kim et al., 30 Sep 2025). This is mathematically distinct from simply conditioning an otherwise isotropic denoiser on key states.

Autoregressive video planning introduces yet another factorization. TempAct models video generation as

$\rightarrow$0

and inserts hierarchy by defining a planner policy $\rightarrow$1 over span-aware step prompts and an executor policy that conditions on smoothed prompt transitions while rolling out latent frames under its own generated histories (Wang et al., 26 Jun 2026).

4. Coupling, feedback, and guidance

A central issue in hierarchical diffusion planning is whether high-level generation remains compatible with low-level execution. CHD attributes long-horizon failure to loose coupling between high-level sub-goal selection and low-level trajectory generation, and addresses it by training a shared classifier that predicts segment reward-to-go and injects classifier gradients into both the high-level and low-level reverse kernels (Hao et al., 12 May 2025). The result is not merely sequential conditioning; it is a joint denoising process in which low-level quality feeds back into subgoal formation during sampling.

TempAct addresses an analogous mismatch in chunk-wise autoregressive video generation. It identifies a sequence of failure modes—temporal instruction ambiguity under a single global prompt, delayed reactions and blended step semantics under naïve step-wise prompting, and error propagation across prompt transitions—and argues that supervised fine-tuning and rollout-based distillation do not directly enforce action ordering and prompt-transition correctness (Wang et al., 26 Jun 2026). Its hierarchical group exploration samples planning groups of candidate plans and, for each plan, execution groups of multiple continuations from a shared visual context. Group-relative advantages are then computed separately for planner-level and executor-level rewards, with GSPO updating the planner and Flow-GRPO updating the executor (Wang et al., 26 Jun 2026).

HD-ExpIt uses a different mechanism for compatibility. Instead of an explicit critic between levels, it alternates supervised retraining with on-policy rollouts and filters for the first successful rollout per context, up to a task-specific cap. Because the high-level planner is retrained only on subgoal sequences from successful trajectories, “HL never sees ‘infeasible’ plans,” and the training distribution is continually trimmed to the low-level controller’s feasible subspace (Grislain et al., 5 Mar 2026).

DOPPLER and XDiffuser illustrate two different placements of search. DOPPLER samples a batch of diffusion-generated options, injects a determinantal repulsion term through $\rightarrow$2 during posterior sampling to increase diversity, and selects the best option according to a learned option-value critic in a receding-horizon loop (Feng et al., 2024). XDiffuser argues that exploration should be extrinsic rather than intrinsic: it first runs Dijkstra or A* on a local-connectivity graph in Temporal Distance Representation space, downsamples the graph path into waypoints, and only then performs a single guided denoising pass (Hassidof et al., 16 May 2026). This establishes a substantive methodological divide within the literature.

5. Empirical performance across domains

On offline reinforcement-learning benchmarks, hierarchical diffusion consistently improves both score and planning efficiency relative to flat diffusion baselines. Hierarchical Diffuser raises Maze2D single-task average performance from 119.5 to 139.9, Multi2D average from 129.4 to 149.9, AntMaze average from 36.0 to 88.8, Gym-MuJoCo average from 77.5 to 84.6, and FrankaKitchen average from 53.1 to 72.5; it also reports up to a $\rightarrow$3 speed-up in planning and effective $\rightarrow$4 speed-up especially on long-horizon tasks (Chen et al., 2024). HM-Diffuser-X, which combines multiscale planning with Progressive Trajectory Extension, reaches 142.1 average score versus 84.0 for Diffuser-X and 113.0 for plain HD-X in single-task Maze2D, 177.1 versus 111.2 and 130.4 in multi-task Maze2D, and 245.4±9.8 versus 98.8±2.7 and 91.6±4.6 in the Giant maze (Chen et al., 25 Mar 2025). CHD reports a further 10–15% gain in normalized reward over flat Diffuser and hierarchical baselines on D4RL Maze2D, completes $\rightarrow$5 of 10 subtasks in Kitchen World versus 5.49 for BHD and 4.09 for Diffuser, and uses 30–40% fewer sampling steps (Hao et al., 12 May 2025).

In motion planning and navigation, the hierarchy frequently separates geometric global structure from locally reactive execution. The pursuit–evasion planner reports up to a 77.2% reduction in detection, 47.4% improvement in goal reaching, and 51.4% increase in overall score relative to the best non-learning heuristic, with all improvements significant at $\rightarrow$6 (Wu et al., 2024). PoPi achieves 10/10, 8/10, and 8/10 success on 2 m, 6 m, and 10 m office-chair goals, compared with 10/10, 6/10, and 5/10 for online RRT, and retains 50–70% success across eight zero-shot generalization conditions involving layout, grasp, chair, and flooring changes (Ravan et al., 2024). Cascaded diffusion models for neural motion planning attain 91.2% success on Turtlebot, 98.0% on Franka Hybrid, and 85.1% on Franka Global, with inference times of 1.25 s and 2.74 s for the latter two settings; a single-pass repair stage guarantees the returned trajectory has zero states in collision (Sharma et al., 21 May 2025).

Structured priors and hybrid generative mechanisms also show large gains. The task-conditioned uncertainty-aware planner achieves 75/100 Maze2D successes after 80 k steps, versus 36/100 for isotropic/no keys, 42/100 for isotropic/with keys, and 14/100 for GPMP/no keys; on KUKA block stacking after 30 k steps it reaches 71/100, versus 11/100, 12/100, and 8/100 for the same baselines (Kim et al., 30 Sep 2025). HDFlow reports simulation success rates of 92 / 71 / 39 on one_leg, 68 / 49 / 34 on lamp, 61 / 43 / 27 on round_table, and 55 / 36 on cabinet, while keeping total replanning cost at approximately 88 ms per planning step on a single GPU; in real-world evaluation it reaches 8/10 and 6/10 on one_leg, 5/10 and 4/10 on lamp, and 4/10 and 3/10 on round_table under low/medium randomization (Gireesh et al., 6 May 2026). XDiffuser reaches 98.5% success on AntMaze-Explore Large versus 27% for CD, 100% on Giant mazes versus 68% for CD, approximately 58% joint success in four-agent planning versus approximately 13% for Prioritized CDGS, and approximately 95–100% POI coverage in TSP-style inspection at $\rightarrow$7, versus approximately 20–33% for CDGS (Hassidof et al., 16 May 2026).

In language-conditioned manipulation and video generation, hierarchical diffusion likewise functions as a long-horizon composition mechanism. SkillDiffuser reports 43.0±1.1% average success on LOReL Sawyer and 23.3% on Meta-World MT10, exceeding the flat visual planner, language-only diffuser, and LISA baselines listed in the paper (Liang et al., 2023). After three iterations, HD-ExpIt reaches 97.5%, 92.7%, 86.6%, 79.3%, and 71.3% success at 1–5 consecutive CALVIN instructions, with AvgLen 4.28; the initial hierarchical diffusion policy at iteration 0 records AvgLen 2.69 (Grislain et al., 5 Mar 2026). TempAct raises average Temporal-Following from 0.400 to 0.462 on Self-Forcing and from 0.432 to 0.488 on LongLive, reports no degradation in VBench quality or PickScore, and obtains more than 80% agreement in a human study on improved temporal consistency (Wang et al., 26 Jun 2026).

6. Limitations, debates, and open directions

The literature identifies several recurring failure modes. TempAct states that supervised fine-tuning suffers from exposure bias and that rollout-based distillation still optimizes low-level denoising or teacher-distribution matching rather than directly enforcing action ordering and prompt-transition correctness (Wang et al., 26 Jun 2026). CHD argues that independent high-level and low-level planning can yield incoherent plans because of loose coupling (Hao et al., 12 May 2025). These critiques are directed not at diffusion itself, but at the interface between abstraction and execution.

A second debate concerns where exploration should occur. XDiffuser argues that intrinsic search inside denoising is expensive, contrasting extrinsic search cost $\rightarrow$8 plus one diffusion pass $\rightarrow$9 with intrinsic search requiring $\tau^{\mathrm{SD}}=[s_0,s_K,\dots,s_{HK}]$0 denoiser calls (Hassidof et al., 16 May 2026). By contrast, CHD and TempAct place feedback directly inside the sampling or update loop rather than outside it (Hao et al., 12 May 2025, Wang et al., 26 Jun 2026). This suggests a methodological split between externally scaffolded diffusion and internally coupled diffusion.

A third issue is extrapolation beyond the support of the training data. HM-Diffuser explicitly states that Diffuser-style planners cannot natively generate trajectories longer than those seen in training, motivating Progressive Trajectory Extension (Chen et al., 25 Mar 2025). The task-conditioned uncertainty-aware planner reports that GPMP priors alone oversmooth and hurt success, while simple conditioning on key states under isotropic noise yields only marginal improvements; the gains arise when structure is embedded in the noise model itself (Kim et al., 30 Sep 2025). A plausible implication is that long-horizon generalization depends not only on hierarchical decomposition, but also on how trajectory priors are represented in the forward corruption process.

Finally, several systems retain classical robotics limitations. PoPi lists no recovery from complete grasp loss, no integrated grasp selection, and reliance on accurate global localization and a prior obstacle map (Ravan et al., 2024). These are concrete reminders that hierarchical diffusion planning, even when successful as a generative planner, may still depend on external state estimation, map quality, contact assumptions, or repair procedures. The field’s recent trajectory—structured priors, coupled feedback, extrinsic graph guidance, multiscale extension, and planner–executor reinforcement learning—suggests that future progress will likely depend on how these auxiliary mechanisms are combined rather than on diffusion denoising alone.