Adaptive Diffusion Policy (ADP) Overview

• Adaptive Diffusion Policy (ADP) is a family of diffusion-based action generation strategies that dynamically adjust conditioning, sampling, and compute allocation according to context.
• Various implementations, such as ADM-DP, ADPro, and RA-DP, leverage multimodal fusion, geometric guidance, and adaptive denoising to improve robotics and reinforcement learning outcomes.
• Empirical studies show ADP variants achieving significant gains in task success, inference speed, and generalization, while addressing challenges like sensor noise and real-world variability.

Searching arXiv for the cited ADP-related papers to ground the article in current records. Adaptive Diffusion Policy (ADP) is a non-unified term used in several recent works to denote diffusion-based action policies that adapt to task context, sensing conditions, domain shifts, or inference-time compute allocation. In the context of multi-agent manipulation, ADP refers to a diffusion-based visuomotor policy that conditions the denoising process on multi-modal observations whose relative importance is adapted online to task context such as approach, contact, and close-proximity coordination; ADM-DP is a concrete instantiation with vision, tactile, and graph fusion (Wang et al., 25 Feb 2026). In a different usage, ADP denotes a training-free test-time adaptation mechanism that constrains reverse diffusion with geometric manifold guidance and analytically guided initialization; ADPro is its implementation atop a pre-trained diffusion policy (Li et al., 8 Aug 2025). Related works use the same broader label for closed-loop adaptation in non-stationary reinforcement learning (Baveja, 31 Mar 2025), zero-shot domain adaptation via lagged context representations and biased diffusion priors (Wang et al., 3 Feb 2026), training-free high-frequency replanning (Ye et al., 6 Mar 2025), adaptive denoising-step allocation (Yu et al., 9 Aug 2025), vision-driven adaptive training and inference budgets (Yu et al., 17 Apr 2026), and adaptive-gradient fine-tuning of diffusion policies in reinforcement learning (Jiang et al., 13 May 2025). This suggests that ADP is best understood as a family of adaptation strategies layered onto diffusion policies rather than a single canonical algorithm.

1. Terminology and scope

The phrase “Adaptive Diffusion Policy” is used across the literature with different operational meanings. In some papers, “adaptive” refers to online re-weighting of input modalities; in others, it refers to training-free guidance during sampling, domain-aware priors, adaptive denoising budgets, or adaptive-gradient optimization. The common element is that diffusion-based action generation is modified so that the reverse process responds to some notion of context, difficulty, geometry, or task phase.

Usage of ADP	Concrete instantiation	Adaptive element
Multi-modal manipulation	ADM-DP (Wang et al., 25 Feb 2026)	AMAM re-weights vision, tactile, and graph modalities online
Test-time geometric adaptation	ADPro (Li et al., 8 Aug 2025)	Chamfer-gradient guidance and Fast Global Registration initialization without retraining
Non-stationary visual RL	Diffuser-based ADP (Baveja, 31 Mar 2025)	Closed-loop resampling and continual fine-tuning
Domain adaptation	DADP (Wang et al., 3 Feb 2026)	Lagged Context Dynamical Prediction and domain-aware diffusion injection
High-frequency replanning	RA-DP (Ye et al., 6 Mar 2025)	Training-free guidance at every denoising step and an action queue
Adaptive compute allocation	D3P (Yu et al., 9 Aug 2025), VADF (Yu et al., 17 Apr 2026), AdaDiff (Zhang et al., 2023)	State-, stage-, or prompt-dependent denoising budgets
Adaptive optimization	ADPO (Jiang et al., 13 May 2025)	AdamW or ADAPG fine-tuning of diffusion policies

A further source of ambiguity is that “diffusion” itself is overloaded. In robotics ADP papers it usually denotes denoising diffusion models for action generation, whereas Erginbas, Vlaski, and Sayed use adaptive combination policies for diffusion learning over networks, where diffusion denotes distributed adaptive learning over graph edges rather than generative denoising (Erginbas et al., 2020). This suggests that the term should be interpreted from the paper-specific formalism rather than from the acronym alone.

2. Shared diffusion-policy substrate

Despite the diversity of meanings, many ADP formulations retain the same basic denoising-policy substrate: actions or action sequences are modeled by a forward noising process and a conditional reverse process. ADM-DP makes this explicit with a $K$-step forward diffusion over clean action trajectories $a^0$,

$$q(a^k \mid a^{k-1}) = \mathcal{N}(a^k; \sqrt{1-\beta_k}\,a^{k-1}, \beta_k \mathbf{I}), \qquad k=1,\ldots,K,$$

and a reverse process parameterized by a U-Net conditioned on multimodal features $f_{\mathrm{cond}}$,

$$p_\theta(a^{k-1} \mid a^k, f_{\mathrm{cond}})=\mathcal{N}(a^{k-1}; \mu_\theta(a^k,k,f_{\mathrm{cond}}), \sigma_k^2 \mathbf{I}),$$

with the standard noise-prediction objective augmented by AMAM entropy regularization (Wang et al., 25 Feb 2026).

Closely related DDPM formulations appear in non-stationary visual reinforcement learning, where action sequences $x$ are diffused and denoised under observation conditioning $\phi(s)$:

$$q(x_t \mid x_{t-1}) = \mathcal{N}(\sqrt{\alpha_t}x_{t-1}, \beta_t I), \qquad p_\theta(x_{t-1} \mid x_t, s)=\mathcal{N}(\mu_\theta(x_t,t,\phi(s)), \Sigma_\theta(t)),$$

with training by a denoising objective $L_{\mathrm{diff}}$ or an equivalent mean-squared error in $\mu$-parameterization (Baveja, 31 Mar 2025). In this family, adaptation is usually expressed not by replacing diffusion itself, but by altering the conditioning variables, the prior, the denoising schedule, or the reverse update.

The control loop is likewise shared but specialized differently. ADM-DP employs DDIM sampling with 20 denoising steps, conditions on the latest 3 observation frames, predicts an action chunk of horizon $a^0$0 timesteps, executes the first 6, and then replans (Wang et al., 25 Feb 2026). RA-DP preserves DDIM-style denoising but reformulates the training setup with independent per-action noise levels so that one executable action can be produced at every denoising step (Ye et al., 6 Mar 2025). D3P retains a DDIM-style sampler but lets a lightweight adaptor choose how many noise levels to skip for each action, turning the denoising budget itself into a policy variable (Yu et al., 9 Aug 2025).

3. Principal modes of adaptation

One major ADP line adapts the conditioning signal. ADM-DP augments Diffusion Policy and DP3 with FiLM-enhanced RGB-point-cloud fusion, tactile-guided grasp refinement, graph attention over tool center point positions, and an Adaptive Modality Attention Mechanism,

$a^0$1

so that modality importance shifts with task phase (Wang et al., 25 Feb 2026). VADF extends the same broad idea into both training and deployment: ALN adaptively samples difficult timesteps and hard trajectories during training, while HVTS decomposes a task into semantic stages and assigns stage-dependent pairs $a^0$2 so that simple stages receive shorter noise schedules and longer execution horizons, whereas complex stages receive longer denoising and shorter horizons (Yu et al., 17 Apr 2026).

A second line adapts the reverse trajectory at test time without retraining. ADPro augments a pre-trained diffusion policy by applying observation-guided Chamfer-distance gradients and a spherical Gaussian step-size constraint to the reverse update,

$a^0$3

and initializes from a structured noisy action computed by Fast Global Registration,

$a^0$4

thereby reducing the remaining denoising horizon and aligning generation with scene geometry (Li et al., 8 Aug 2025). RA-DP also operates training-free at inference time, but its mechanism is different: it injects a differentiable guidance loss $a^0$5 at every denoising step by a normalized Polyak-style update in latent action space, while an action queue yields one executable action per denoising step (Ye et al., 6 Mar 2025).

A third line adapts the prior and denoising target to latent domain information. DADP learns a static domain representation $a^0$6 with Lagged Context Dynamical Prediction, using a historical offset context

$a^0$7

so that domain representations filter out transient properties. It then injects $a^0$8 directly into the diffusion process through a biased prior,

$a^0$9

and a reformulated diffusion target

$$q(a^k \mid a^{k-1}) = \mathcal{N}(a^k; \sqrt{1-\beta_k}\,a^{k-1}, \beta_k \mathbf{I}), \qquad k=1,\ldots,K,$$0

so that denoising follows the domain-specific action manifold more directly (Wang et al., 3 Feb 2026).

A fourth line adapts the amount of computation. D3P treats denoising-step allocation as a reinforcement-learning problem in a two-layer partially observable Markov decision process, where a lightweight adaptor $$q(a^k \mid a^{k-1}) = \mathcal{N}(a^k; \sqrt{1-\beta_k}\,a^{k-1}, \beta_k \mathbf{I}), \qquad k=1,\ldots,K,$$1 observes $$q(a^k \mid a^{k-1}) = \mathcal{N}(a^k; \sqrt{1-\beta_k}\,a^{k-1}, \beta_k \mathbf{I}), \qquad k=1,\ldots,K,$$2 and chooses a stride $$q(a^k \mid a^{k-1}) = \mathcal{N}(a^k; \sqrt{1-\beta_k}\,a^{k-1}, \beta_k \mathbf{I}), \qquad k=1,\ldots,K,$$3 for the current noisy action chunk, allowing crucial actions to receive more denoising and routine actions fewer (Yu et al., 9 Aug 2025). AdaDiff, although developed for image and video generation rather than robotic control, follows the same principle at the prompt level: a learned policy selects one step count from $$q(a^k \mid a^{k-1}) = \mathcal{N}(a^k; \sqrt{1-\beta_k}\,a^{k-1}, \beta_k \mathbf{I}), \qquad k=1,\ldots,K,$$4 before sampling, balancing quality and efficiency (Zhang et al., 2023). ADPO shifts adaptation to the optimizer, using AdamW or the proposed ADAPG update,

$$q(a^k \mid a^{k-1}) = \mathcal{N}(a^k; \sqrt{1-\beta_k}\,a^{k-1}, \beta_k \mathbf{I}), \qquad k=1,\ldots,K,$$5

to stabilize and accelerate reinforcement-learning fine-tuning of diffusion policies (Jiang et al., 13 May 2025).

4. ADM-DP as a concrete multi-agent ADP

ADM-DP provides one of the most explicit and operational definitions of ADP for robotics. Each agent’s observation is decomposed into local and shared components,

$$q(a^k \mid a^{k-1}) = \mathcal{N}(a^k; \sqrt{1-\beta_k}\,a^{k-1}, \beta_k \mathbf{I}), \qquad k=1,\ldots,K,$$6

where

$$q(a^k \mid a^{k-1}) = \mathcal{N}(a^k; \sqrt{1-\beta_k}\,a^{k-1}, \beta_k \mathbf{I}), \qquad k=1,\ldots,K,$$7

contains RGB image $$q(a^k \mid a^{k-1}) = \mathcal{N}(a^k; \sqrt{1-\beta_k}\,a^{k-1}, \beta_k \mathbf{I}), \qquad k=1,\ldots,K,$$8, point cloud $$q(a^k \mid a^{k-1}) = \mathcal{N}(a^k; \sqrt{1-\beta_k}\,a^{k-1}, \beta_k \mathbf{I}), \qquad k=1,\ldots,K,$$9, tactile readings $f_{\mathrm{cond}}$0, joint states $f_{\mathrm{cond}}$1, and a language instruction $f_{\mathrm{cond}}$2, while the shared observation is the set of all agents’ TCP positions,

$f_{\mathrm{cond}}$3

Encoders map $f_{\mathrm{cond}}$4 to a conditioning vector $f_{\mathrm{cond}}$5 that drives the diffusion model (Wang et al., 25 Feb 2026).

The visual pathway combines image semantics and 3D geometry. RGB images are processed by a ResNet encoder to produce $f_{\mathrm{cond}}$6, while point clouds are cropped to the workspace, downsampled to $f_{\mathrm{cond}}$7 points with retained color, and processed by a modified PointNet that removes T-Net and BatchNorm following DP3 to produce $f_{\mathrm{cond}}$8. FiLM uses point-cloud geometry to modulate image semantics,

$f_{\mathrm{cond}}$9

with the stated aim of geometry-aware modulation of semantics for robust 3D understanding under occlusions.

The tactile pathway is designed around 4×4 FSRs per fingertip, giving 32 readings per gripper. Inputs are reshaped to $$p_\theta(a^{k-1} \mid a^k, f_{\mathrm{cond}})=\mathcal{N}(a^{k-1}; \mu_\theta(a^k,k,f_{\mathrm{cond}}), \sigma_k^2 \mathbf{I}),$$0, log-normalized, and concatenated with positional encodings $$p_\theta(a^{k-1} \mid a^k, f_{\mathrm{cond}})=\mathcal{N}(a^{k-1}; \mu_\theta(a^k,k,f_{\mathrm{cond}}), \sigma_k^2 \mathbf{I}),$$1 per taxel. Per finger, 1D convolutions with adaptive pooling extract spatial features, and physical features are computed as

$$p_\theta(a^{k-1} \mid a^k, f_{\mathrm{cond}})=\mathcal{N}(a^{k-1}; \mu_\theta(a^k,k,f_{\mathrm{cond}}), \sigma_k^2 \mathbf{I}),$$2

These are fused by an MLP to yield $$p_\theta(a^{k-1} \mid a^k, f_{\mathrm{cond}})=\mathcal{N}(a^{k-1}; \mu_\theta(a^k,k,f_{\mathrm{cond}}), \sigma_k^2 \mathbf{I}),$$3. No fixed thresholds are hard-coded for contact insufficiency. Instead, demonstrations include 30% shallow-to-deep refinements, allowing the policy to learn tactile signatures of insufficient contact and to execute either release–regrasp or in-grasp tightening when weak or incomplete contact is detected.

The graph pathway represents multi-agent collision context through a fully connected graph over TCP positions with distance-based edge weights

$$p_\theta(a^{k-1} \mid a^k, f_{\mathrm{cond}})=\mathcal{N}(a^{k-1}; \mu_\theta(a^k,k,f_{\mathrm{cond}}), \sigma_k^2 \mathbf{I}),$$4

A standard Graph Attention Network produces $$p_\theta(a^{k-1} \mid a^k, f_{\mathrm{cond}})=\mathcal{N}(a^{k-1}; \mu_\theta(a^k,k,f_{\mathrm{cond}}), \sigma_k^2 \mathbf{I}),$$5, focusing coordination at end-effectors rather than along full kinematic chains. AMAM then computes context-aware modality weights, and the fused representation is concatenated with joint states and FiLM-modulated by a frozen CLIP text encoder:

$$p_\theta(a^{k-1} \mid a^k, f_{\mathrm{cond}})=\mathcal{N}(a^{k-1}; \mu_\theta(a^k,k,f_{\mathrm{cond}}), \sigma_k^2 \mathbf{I}),$$6

ADM-DP adopts a decoupled training paradigm in which each agent learns an independent policy $$p_\theta(a^{k-1} \mid a^k, f_{\mathrm{cond}})=\mathcal{N}(a^{k-1}; \mu_\theta(a^k,k,f_{\mathrm{cond}}), \sigma_k^2 \mathbf{I}),$$7 while receiving the shared TCP set during training and inference. The training loop consists of demonstration collection with motion planners, multimodal encoding, adaptive fusion, FiLM conditioning with language, and diffusion training with

$$p_\theta(a^{k-1} \mid a^k, f_{\mathrm{cond}})=\mathcal{N}(a^{k-1}; \mu_\theta(a^k,k,f_{\mathrm{cond}}), \sigma_k^2 \mathbf{I}),$$8

There are no explicit collision or grasp-stability penalty losses. Collision avoidance is achieved through proximity-sensitive graph conditioning, and grasp refinement is learned from demonstrations. At inference, all agents run in parallel, recompute AMAM weights online each cycle, sample via DDIM, execute the first 6 actions of an $$p_\theta(a^{k-1} \mid a^k, f_{\mathrm{cond}})=\mathcal{N}(a^{k-1}; \mu_\theta(a^k,k,f_{\mathrm{cond}}), \sigma_k^2 \mathbf{I}),$$9 chunk, and replan on the next observation cycle.

5. Empirical evidence across ADP variants

Reported results support the practical utility of adaptive mechanisms, but they do not reduce to a single benchmark narrative. Some papers emphasize success-rate gains from better sensing and coordination, others emphasize faster inference, earlier success, or improved generalization under unseen geometry and dynamics.

Paper	Setting	Reported outcome
ADM-DP (Wang et al., 25 Feb 2026)	Seven multi-agent manipulation tasks	50 demos: 37.6% vs best baseline 24.3%; 150 demos: 57.1% vs 45.4%; gains of 12–25%
ADPro (Li et al., 8 Aug 2025)	RLBench, CALVIN, RealWP, Acronym	Up to 25% faster execution and 9% points over strong diffusion baselines; CALVIN average sequence length 3.64 vs 3.35 and execution time 36.2 s vs 48.3 s
RA-DP (Ye et al., 6 Mar 2025)	MetaWorld and real Franka replanning	Point-cloud MetaWorld average 68.6 vs 58.3 and RF 82.3 Hz vs 16.1 Hz; static obstacle 63.3% vs 15.0%; dynamic obstacle 45.0% vs 10.0%
D3P (Yu et al., 9 Aug 2025)	Robomimic, Franka Kitchen, real Franka	Averaged 2.2× inference speed-up in simulation and 1.9× on a physical robot; 33.68 Hz vs 17.59 Hz on real Square
VADF (Yu et al., 17 Apr 2026)	Robomimic, Kitchen, Adroit, ARX5	DP-C 0.92 → 0.95; DP-T 0.84 → 0.88; early success +5.9% and +6.7%; DDIM+VADF up to 2.46× speedup
AdaDiff (Zhang et al., 2023)	Image and video generation	At least 33% and up to 40% inference-time reduction with similar quality; 50.8% speedup with DPM-Solver on COCO 2017

Within ADM-DP, task-level and ablation results localize where adaptation matters most. Across seven tasks with Franka Panda arms, 4×4 FSR per fingertip, and 50 or 150 demonstrations per agent, the strongest gains are reported in tasks where tactile and graph cues are both relevant. At 150 demonstrations, Lift Barrier reaches 92% versus 77% for DP3, Lift Arm 78% versus 62% for DP3, and Pass Peg 74% versus 53% for Flow Policy. Removing point clouds drops the average to 45.9% and lowers Two Robots Stack Cube from 44% to 26%; removing tactile reduces the average to 50.9% and lowers Lift Barrier from 92% to 78%; removing graph reduces the average to 53.7%, including a 7-point drop on Pass Peg and a 9-point drop on Pass Shoe; removing AMAM reduces the average to 52.0%, with the largest drops on multimodal tasks such as Pass Shoe (Wang et al., 25 Feb 2026).

ADPro’s ablations identify observation guidance as the most critical component. Removing the initial noise constraint lowers average success from 83.9 to 82.4 and increases time; removing the spherical Gaussian constraint lowers success to 82.1; removing observation guidance causes the largest drop to 81.6 and raises time from 435.8 s to 488.7 s (Li et al., 8 Aug 2025). DADP reports that across Ant, HalfCheetah, Walker2d, Hopper, and Adroit tasks it consistently matches or outperforms strong baselines such as CORRO, Prompt-DT, Meta-DT, and DV in both IID and OOD settings, with reduced variance across seeds, and that larger lag improves representation quality markedly in Walker2d and HalfCheetah (Wang et al., 3 Feb 2026).

The non-stationary reinforcement-learning study provides an important counterpoint. There, the diffuser-based ADP achieves mean 8.15, max 8.30, std 0.15 on CoinRun, outperforming PPO and DQN on mean and max reward; and in PointMaze it achieves the highest maximum reward, 98.50, with lower variability than PPO and DQN. However, in Maze it underperforms both PPO and DQN, with mean 3.23 compared with PPO’s 8.70 and DQN’s 8.20 (Baveja, 31 Mar 2025). This indicates that adaptive diffusion policies do not uniformly dominate simpler model-free baselines across all task structures.

6. Limitations, misconceptions, and open directions

A common misconception is that ADP denotes a single established algorithm. The literature instead uses the term for several distinct interventions: adaptive modality fusion in ADM-DP, training-free geometry-aware reverse diffusion in ADPro, continual closed-loop adaptation in non-stationary RL, domain-conditioned prior shaping in DADP, training-free replanning in RA-DP, denoising-budget allocation in D3P and VADF, and adaptive-gradient optimization in ADPO. This suggests that the most stable definition of ADP is functional rather than architectural: a diffusion policy is “adaptive” when some component of conditioning, sampling, prior selection, compute allocation, or optimization changes with context.

A second misconception is that stronger adaptation necessarily requires explicit penalties or monolithic joint policies. ADM-DP explicitly reports that there is no collision penalty term in the loss and no explicit grasp-stability penalty loss; coordination is achieved through graph-conditioned multimodal diffusion, and grasp refinement is learned from demonstrations (Wang et al., 25 Feb 2026). Likewise, RA-DP and ADPro are training-free at deployment and adapt through guidance signals or geometric priors rather than parameter updates (Ye et al., 6 Mar 2025, Li et al., 8 Aug 2025).

The limitations are correspondingly heterogeneous. ADM-DP is evaluated in simulation, with future work aimed at real robot deployment under sensor noise and real-time constraints, and its TCP-based graph may require enhanced coordination mechanisms beyond three robots (Wang et al., 25 Feb 2026). ADPro depends on test-time gripper and scene point clouds and can degrade under occlusions, sensor noise, or dynamic targets (Li et al., 8 Aug 2025). DADP assumes stationary dynamics within an episode, so non-stationary domains where static factors evolve would violate its disentanglement premise (Wang et al., 3 Feb 2026). RA-DP can fail when environment changes outpace replanning frequency, with success approaching 0 beyond a normalized dynamic-obstacle speed of 0.14 (Ye et al., 6 Mar 2025). D3P is sensitive to adaptor misprediction and to hyperparameters such as $x$0, $x$1, $x$2, $x$3, and $x$4 (Yu et al., 9 Aug 2025). VADF depends on zero-shot VLM stage recognition and can exhibit jitter near stage boundaries or suboptimal schedule assignment when stage cues are poorly visible (Yu et al., 17 Apr 2026). ADPO reports that $x$5 is robust but that the Katyusha-style momentum coefficient $x$6 remains task- and method-dependent (Jiang et al., 13 May 2025).

An additional terminological caution arises outside robotic denoising policies. In distributed learning over networks, Erginbas, Vlaski, and Sayed derive Gramian-based adaptive combination policies that adjust edge weights by solving a constrained quadratic program based on a centered Gramian of intermediate estimates, thereby improving transient learning performance while preserving the steady-state benefits of variance-based rules (Erginbas et al., 2020). This is an adaptive diffusion policy in a different sense: the adaptive object is the network combination matrix rather than a generative reverse-diffusion controller.

Across these formulations, the durable technical theme is not one fixed recipe but a recurrent design pattern: diffusion-based action generation is made contingent on information that a non-adaptive policy would treat as static. In different papers, that contingency is supplied by multimodal attention, geometric losses, lagged domain context, stage segmentation, stride selection, or optimizer state. The research direction therefore remains open-ended, with future work already pointing toward real-robot deployment, richer tactile sensing, more agents, denser workspaces, uncertainty-aware compute allocation, and more robust adaptation under sensing noise and distribution shift.