---
title: Query-centric Diffusion Policy for Robotic Assembly
url: https://www.emergentmind.com/topics/query-centric-diffusion-policy-qdp
type: topic
---

# Query-centric Diffusion Policy for Robotic Assembly

Query-centric Diffusion Policy (QDP) is a hierarchical framework for robotic assembly that bridges high-level planning and low-level control by utilizing queries comprising objects, contact points, and skill information. It is designed for contact-rich assembly settings in which part interactions are intrinsically complex and execution is sensitive to noise perturbations. In QDP, the query is the central interface between semantic task decomposition and geometry-aware action generation: a high-level policy produces the next skill, object, and contact point of focus, and a low-level diffusion policy uses that query to guide perception and generate operational-space action chunks from point cloud observations and robot proprioception. On FurnitureBench in simulation and real-world settings, QDP is reported to improve skill precision and long-horizon success rate; in the challenging insertion and screwing tasks, it improves the skill-wise success rate by over 50% compared to baselines without structured queries [2509.18686].

## 1. Problem Setting and Motivation

Robotic assembly is typically organized hierarchically, with high-level multi-part reasoning separated from low-level precise control. In practice, this decomposition is difficult because high-level skill queries are often abstract or ambiguous, whereas low-level execution in contact-rich tasks such as insertion and screwing requires highly specific 6-DoF end-effector trajectories. The central motivation of QDP is to reduce this mismatch by making the query express “what to do, to which, where,” so that planning, perception, and control are all conditioned on the same structured specification [2509.18686].

The target domain in the reported study is furniture assembly, where the agent must reason over multiple parts, identify relevant contact locations, and maintain robustness under clutter, perturbations, and sim-to-real discrepancies. This setting makes the semantic-to-physical gap especially acute: a planner may know that a leg must be inserted into a hole, but execution still depends on local geometry, pose accuracy, and contact alignment. QDP addresses this by explicitly encoding skill, part, and contact point in a single query and using that query continuously rather than only at skill boundaries.

A plausible implication is that QDP treats hierarchical control not merely as temporal abstraction, but as a representational alignment problem. The framework does not only decompose the task into high-level and low-level stages; it defines a common symbolic-geometric interface through which both stages communicate.

## 2. Hierarchical Framework

QDP is a two-level hierarchical policy consisting of a high-level policy, described as a “Sequence Generator,” and a low-level policy, described as a “Query-conditioned Action Generator.” The high-level policy takes as input the initial and current RGB scene image, augmented with segmentation masks and Set-of-Mark (SoM) markers for visual grounding. Its core engine is Vision-Language Model Predictive Control (VLM-PC), enhanced with object and contact-point recognition using powerful vision-language models such as GPT-4o. At each step, VLM-PC uses the current image and interaction history to dynamically select the next skill, object, and contact point of focus, and it replans after each skill execution [2509.18686].

The high-level output is a query that specifies three components: the skill $\ell$ (for example, grasp, insert, or screw), the object or furniture part $i$, and the contact point $j$. This query is then consumed by the low-level policy. The low-level policy takes the query $q_{ij}^{(\ell)}$, a workspace point cloud, and current robot proprioception. A PointNet-based encoder conditioned on the query estimates the pose of the specific part and contact point in the point cloud. A diffusion-model-based action decoder, conditioned on the query and state, produces a precise sequence, or “chunk,” of end-effector actions in operational space, and the output is tracked by an operational space controller (OSC) [2509.18686].

This decomposition is significant because the query is not a loose textual prompt or a latent code; it is an explicit structured variable shared across levels. The reported interpretation is that this avoids confusion from irrelevant context such as background geometry or uninvolved parts, thereby bridging high-level planning and low-level control hierarchically and continually.

## 3. Query Formalization

The core formal object in QDP is the query set
$$
\mathcal{Q} \triangleq \Bigl\{\,q=(q_{ij}^{(\ell)})\in\{0,1\}^{N_s\times N_p\times N_c} \;\Bigm|\; \sum_{\ell=1}^{N_s}\sum_{i=1}^{N_p}\sum_{j=1}^{N_c} q_{ij}^{(\ell)} =1 \Bigr\}.
$$
Here, $N_s$ is the number of skill types, $N_p$ is the number of furniture parts, and $N_c$ is the number of contact points. At each timestep, a one-hot query $q_{ij}^{(\ell)}$ is generated, selecting exactly one tuple of skill, part, and contact point [2509.18686].

In the reported examples, the skill space includes stabilize, grasp, insert, and screw. A query such as $q_{leg3, hole2}^{(\text{insert})}=1$ encodes the instruction “insert leg 3 into hole 2.” This is the signal passed from the high-level planner to the low-level policy.

The formalization matters because it converts hierarchical planning into a discrete, interpretable selection problem. The query is described as a compact and interpretable signal encoding task intention through the product space of object, contact, and skill. The paper further characterizes it as the hinge between symbolic planning from VLM-PC and geometry/physics-aware control from the diffusion policy. This suggests that QDP’s main contribution is not only a diffusion-based controller, but a specific task interface that makes diffusion control usable for long-horizon assembly.

## 4. Query-Centric Perception and Diffusion Action Generation

The query-centric encoder combines PointNet with embedding layers for skill, part, and contact point, together with proprioception. Given the workspace point cloud and the query, its goal is to estimate only the relevant furniture part’s pose, rotation, and contact point. End-effector state is included to ground predictions in the current manipulator configuration. The reported estimation loss is
$$
L_{\text{EST} = \sum_{q_{ij}^{(\ell)} \in \mathcal Q} q_{ij}^{(\ell)} \cdot \big( \alpha   \| \hat{P}_i - P_i \|^2 + \beta  \| \hat{R}_i - R_i \|^2 + \gamma \| \hat{P}_j - P_j \|^2 \big)
$$
with $\hat{P}_i$, $\hat{R}_i$, and $\hat{P}_j$ denoting the predicted position, rotation of part $i$, and position of contact $j$, and $P_i$, $R_i$, and $P_j$ the corresponding ground truths. For furniture such as table legs that have axial symmetry, rotation invariance in the $z$-axis is assumed for robustness [2509.18686].

The action decoder is a Temporal CNN-based U-Net with FiLM conditioning. Its inputs are the true state of the queried part, the contact point, and current end-effector proprioception. At test time, action chunks are generated through $K$ denoising steps starting from Gaussian noise:
$$
\mathbf{a}^{k-1} = \alpha_k \left ( \mathbf{a}^k - \gamma_k \sum_{i=1}^{N_p} \sum_{j=1}^{N_c} q_{ij}^{(\ell)} \cdot \mathbf{\epsilon}_\theta (\mathbf{a}^k, k, \mathbf s) \right ) + \sigma_k \epsilon
$$
where $\mathbf{a}^k$ is the action sample at step $k$, $\mathbf{s} = [s_i, s_j, s_e]$ is the concatenated state vector, and the one-hot query ensures that only the relevant skill, part, and contact prediction is used. The reported denoising loss is
$$
L_{\text{ACT} = \left\| \epsilon^k - \sum_{q_{ij}^{(\ell)} \in \mathcal Q} q_{ij}^{(\ell)} \cdot \mathbf{\epsilon}_{\theta}^{(\ell)} (\mathbf{a}^0+\epsilon^k, k, \mathbf{s} )  \right\|^2.
$$

Within the framework, the query does more than provide a conditioning token; it gates both the estimation target and the denoising path. This suggests a form of structured sparsification in which only the state variables relevant to the selected skill-part-contact tuple are exposed to the low-level policy.

## 5. Data, Training, and Sim-to-Real Design

The reported training data consist of 300 full-task simulated trajectories for a square table, with randomized assembly order and state jitter, annotated at every step with queries, part states, and contact points. To address the sim-to-real gap, the system uses synthetic point clouds derived from meshes for the furniture, but real-world point clouds for the gripper [2509.18686].

Training proceeds in two stages. First, the query-centric encoder is trained to estimate poses given a query and a point cloud. Second, skill-specific action decoders are trained conditioned on the query and estimated pose. This separation of estimation and action generation preserves the hierarchical structure inside the low-level policy itself.

The design choice is important for the reported sim-to-real transfer argument. The point-cloud representation is not purely photometric, and the encoder is query-conditioned rather than scene-global. A plausible implication is that the model is encouraged to learn local geometry around the currently relevant part-contact relation rather than a monolithic scene representation, which is consistent with the reported robustness to noisy real-world point cloud data.

## 6. Experimental Results, Ablations, and Failure Modes

QDP is evaluated on FurnitureBench using the Square Table task, in which the robot sequentially assembles four table legs to a tabletop via stabilize, grasp, insert, and screw skills, and on a One Leg Task used for per-skill evaluation. The baselines are Behavior Cloning (BC), DP3, and QDP-A, described as QDP without the query-centric structure [2509.18686].

On low-level per-skill success rate, the reported values are: BC achieves 0.92 on stabilize, 0.41 on grasp, 0.00 on insert, and 0.08 on screw; DP3 achieves 0.83, 0.47, 0.09, and 0.17; QDP achieves 0.95, 0.64, 0.59, and 0.80. The ablation “QDP w/o query” achieves 0.96 on stabilize, 0.60 on grasp, 0.06 on insert, and 0.33 on screw. The paper emphasizes that insert and screw, the most contact-sensitive and precision-demanding skills, see over 50% improvement over the best baseline, and that removing the query-centric structure leads to near-total failure on insert [2509.18686].

The framework is also reported to be robust to shape transformations of parts, randomization of assembly order, and tabletop perturbations. In real-world deployment, QDP is run zero-shot on a Kinova Gen3 robot using four RealSense D435 point clouds. On the One Leg Task, the reported success rate is 7/18. Visualizations are described as showing recovery from disturbances, including human perturbation of the tabletop. Failure analysis states that most failures in the standard and ablated variants occur at precise pose estimation or manipulation, especially during insert or grasp.

Ablations on the high-level policy show that removing VLM-PC, SoM markers, or history causes substantial degradation. The most severe impact comes from removing history, and performance is reported to increase with history window $n$ and plateau after $n \geq 3$ steps. These results position historical context as a necessary component of sequential assembly rather than an optional augmentation.

## 7. Relation to Other Diffusion-Policy Research and Terminological Ambiguity

The abbreviation “QDP” can be ambiguous in the broader diffusion-policy literature. In the robotic assembly paper, QDP denotes Query-centric Diffusion Policy, where the query consists of skill, object, and contact point and is used to bridge high-level planning and low-level execution [2509.18686]. In contrast, the online diffusion-policy RL literature discusses policy-improvement mechanisms such as Action-Gradient, Q-Weighting, Proximity-Based, and Backpropagation Through Time (BPTT), with Q-weighting methods using Q-values or advantages to weight denoising losses rather than structured skill-part-contact queries [2601.06133].

This distinction is particularly important because Q-weighted methods such as Q-weighted Variational Policy Optimization (QVPO) optimize a diffusion model through a Q-weighted variational loss, with weights derived from transformed Q-values or advantages and with explicit discussion of reward-scale sensitivity, sample efficiency, and diffusion-step scalability [2405.16173]. Likewise, Q-score matching links the score of a diffusion policy to the action gradient of a learned Q-function in off-policy RL, again using the letter “Q” in the critic sense rather than in the query-centric sense [2312.11752].

QDP is also related, at a higher level of architectural design, to later hierarchical systems such as WorldDP, which use a high-level world model to optimize feasible subgoals during runtime and a low-level Diffusion Policy to execute them. WorldDP conditions its low-level policy on current and desired future object-centric states, whereas QDP uses a one-hot query over skill, part, and contact point together with point-cloud estimation [2606.08775]. This suggests that QDP belongs to a broader class of hierarchical diffusion-policy frameworks for long-horizon robotics, while remaining distinctive in its discrete query formalization and its explicit focus on contact-rich assembly.

Source: https://www.emergentmind.com/topics/query-centric-diffusion-policy-qdp