GAIDE: Graph-based Attention Masking for Spatial- and Embodiment-aware Motion Planning

Davood Soleymanzadeh; Xiao Liang; Minghui Zheng

cs.ROMar 3, 2026

GAIDE: Graph-based Attention Masking for Spatial- and Embodiment-aware Motion Planning

Authors: Davood Soleymanzadeh, Xiao Liang, Minghui Zheng

Organizations: J. Mike Walker ’66 Department of Mechanical Engineering, Texas A&M University, College Station, TX 77843, USA · Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX 77843 USA

Abstract

Sampling-based motion planning algorithms are widely used for motion planning of robotic manipulators, but they often struggle with sample inefficiency in high-dimensional configuration spaces due to their reliance on uniform or hand-crafted informed sampling primitives. Neural informed samplers address this limitation by learning the sampling distribution from prior planning experience to guide the motion planner towards planning goal. However, existing approaches often struggle to encode the spatial structure inherent in motion planning problems. To address this limitation, we introduce Graph-based Attention Masking for Spatial- and Embodiment-aware Motion Planning (GAIDE), a neural informed sampler that leverages both the spatial structure of the planning problem and the robotic manipulator's embodiment to guide the planning algorithm. GAIDE represents these structures as a graph and integrates it into a transformer-based neural sampler through attention masking. We evaluate GAIDE against baseline state-of-the-art sampling-based planners using uniform sampling, hand-crafted informed sampling, and neural informed sampling primitives. Evaluation results demonstrate that GAIDE improves planning efficiency and success rate.

Figures & tables

Fig. 1: An overview of GAIDE : the proposed neural informed sampler constructs a graph that represents both the robotic manipulator embodiment and the spatial relationships inherent within the motion planning problem. The adjacency matrix of this graph is incorporated into a transformer-based neural sampler via attention masking to enable spatial- and embodiment-aware informed sampling. The framework is embedded within sampling-based planners as an informed sampler and demonstrates superior performance compared to benchmark neural samplers, MPNets [ 6 ] and SIMPNet [ 5 ] .

Fig. 2: Graph Construction : an illustration of spatial and embodiment graph. An undirected graph is constructed over the downsampled manipulator point cloud to implicitly encode the manipulator’s kinematics chain, while a directed graph connects the downsampled workspace point cloud to the robot nodes to capture the inherent spatial relationships within the motion planning problem. “PCD” denotes point cloud and “PointNet++” is set abstraction layers from PointNet++ [ 7 ] .

Fig. 3: GAIDE network architecture: The framework leverages current-time-step workspace information (including the workspace and robot point clouds) together with configuration space features (current-time-step and goal configuration) to generate an informed sample that guides the robotic manipulator toward the motion planning goal. $\mathbf{q}_{t},~\mathcal{P}_{r}$ , and $\mathbf{q}_{\text{goal}}$ are the current time-step configuration, current-time step robotic manipulator point cloud, and motion planning goal configuration, respectively. $\mathbf{z}_{t}$ , $\mathbf{z}_{\text{goal}}$ , $\mathbf{z}_{r}$ , $\mathbf{z}_{w}$ , and $\delta\mathbf{q}_{t}$ are the current time-step configuration, planning goal configuration, robot point cloud, scene point cloud embeddings, and predicted joint angle, respectively. “PointNet++” denotes set abstraction layers from PoinNet++ [ 7 ] .

Fig. 4: An example of all held-out planning tasks from [ 16 ] .

	TableTop		Box		Bins		Shelf
							Task I		Task II		Task III
	T $[s]\downarrow$	S $[\%]\uparrow$	T $[s]\downarrow$	S $[\%]\uparrow$	T $[s]\downarrow$	S $[\%]\uparrow$	T $[s]\downarrow$	S $[\%]\uparrow$	T $[s]\downarrow$	S $[\%]\uparrow$	T $[s]\downarrow$	S $[\%]\uparrow$
Bi-RRT [ 24 ]	$2.85{\pm}0.97$	$\mathbf{88}$ %	$\mathbf{0.73{\pm}0.54}$	$\mathbf{71}$ %	$\mathbf{1.1{\pm}0.67}$	$\mathbf{98.25}$ %	$2.71{\pm}0.73$	$\mathbf{98}$ %	$3.91{\pm}1.12$	$\mathbf{82.6}$ %	$4.07{\pm}1.17$	$\mathbf{45}$ %
RRT* [ 25 ]	$4.15{\pm}0.99$	$50$ %	$2.59{\pm}0.07$	$55$ %	$3.08{\pm}0.1$	$64$ %	$4.29{\pm}1.14$	$51$ %	$5.78{\pm}0.85$	$40.6$ %	$5.12{\pm}1.39$	$29$ %
IRRT* [ 26 ]	$4.44{\pm}1.35$	$48$ %	$2.54{\pm}0.8$	$55$ %	$3.12{\pm}0.09$	$66.5$ %	$4.74{\pm}1.89$	$55$ %	$5.94{\pm}1.04$	$37$ %	$6.65{\pm}1.69$	$27$ %
BIT* [ 27 ]	$3.19{\pm}2.67$	$71$ %	$1.4{\pm}1.47$	$67$ %	$1.84{\pm}1.81$	$93.75$ %	$4.35{\pm}2.38$	$91$ %	$5.95{\pm}2.09$	$75$ %	$4.84{\pm}3.24$	$38$ %

TABLE I: Planning performance of GAIDE and baseline planners across all held-out planning task environments.

Fig. 5: Planning cost of GAIDE and baseline planners across all held-out planning tasks.

Fig. 6: Real-world deployment of GAIDE . The path profile from the start configuration to the goal configuration is demonstrated in the real-world (top), and simulated scene point cloud (bottom) for each demo.

Fig. 7: Average success rate and planning time comparison between GAIDE and benchmark planners across all held-out planning tasks.

	TableTop		Box		Bins		Shelf
							Task I		Task II		Task III
	T $[s]\downarrow$	S $[\%]\uparrow$	T $[s]\downarrow$	S $[\%]\uparrow$	T $[s]\downarrow$	S $[\%]\uparrow$	T $[s]\downarrow$	S $[\%]\uparrow$	T $[s]\downarrow$	S $[\%]\uparrow$	T $[s]\downarrow$	S $[\%]\uparrow$
GAIDE-V	$2.96{\pm}2.71$	$45$ %	$2.15{\pm}1.51$	$62$ %	$3.61{\pm}3.58$	$81.25$ %	$3.26{\pm}1.56$	$45$ %	$\mathbf{3.58{\pm}2.79}$	$36$ %	$\mathbf{4.03{\pm}3.42}$	$33$ %
GAIDE-H	$\mathbf{2.61{\pm}2.71}$	$48$ %	$\mathbf{2.01{\pm}1.25}$	$59$ %	$\mathbf{3.41{\pm}3.04}$	$77.25$ %	$\mathbf{2.48{\pm}1.51}$	$45$ %	$4.58{\pm}2.29$	$34$ %	$4.42{\pm}4.98$	$30$ %
GAIDE	$3.0{\pm}2.58$	$\mathbf{52}$ %	$2.17{\pm}2.02$	$\mathbf{65}$ %	$3.72{\pm}3.56$	$\mathbf{96}$ %	$2.99{\pm}1.58$	$\mathbf{55}$ %	$5.56{\pm}2.57$	$\mathbf{44}$ %	$4.34{\pm}3.25$	$\mathbf{38}$ %

TABLE II: Planning performance of GAIDE and different frameworks that apply attention masking differently across held-out planning environments. “GAIDE-V” denotes GAIDE-Vanilla Transformer, and “GAIDE-H” denotes GAIDE-Hard.

Robot Motion Planning

Robot motion planning finds safe, feasible paths and trajectories for robots moving through constrained, changing environments.

See all 271 papers

May 24, 2026

Performance Comparison of Classical and Neural Sampling Algorithms for Robotic Navigation

Integrating artificial intelligence (AI) into sampling-based motion planning provides new possibilities for improving autonomous navigation efficiency. In this paper, three algorithms, namely RRT*, Neural RRT*, and Neural Informed RRT*, are implemented and evaluated on environments containing convex and concave obstacles with different obstacle densities. The obtained results indicate that neural-guided planners improve path quality, producing up to 14% shorter paths and 55--75% smoother trajectories compared with the conventional RRT* algorithm. Among the evaluated methods, Neural Informed RRT* achieves the best overall performance in terms of path length and trajectory smoothness. These results demonstrate the effectiveness of AI-guided sampling strategies for improving reliability and trajectory efficiency in robotic and UAV navigation, despite a slight increase in computation time. Overall, the study highlights the growing importance of artificial intelligence in real-time robotic path planning applications.

Hichem Cheriet, Badra Khellat Kihel, Samira Chouraqui

dept. of Computer Science Universit´e d’USTO Mohamed Boudiaf Oran, Algeria · dept. of Economics Oran2 Mohamed BenAhmed University Oran, Algeria

Robot Navigation Sampling-Based Motion Planning

Ask PDF

May 6, 2024

Greedy Heuristics for Sampling-Based Motion Planning in High-Dimensional State Spaces

Informed sampling techniques accelerate the convergence of sampling-based motion planners by biasing sampling toward regions of the state space that are most likely to yield better solutions. However, when the current solution path contains redundant or tortuous segments, the resulting informed subset may remain unnecessarily large, slowing convergence. Our prior work addressed this issue by introducing the greedy informed set, which reduces the sampling region based on the maximum heuristic cost along the current solution path. In this article, we formally characterize the behavior of the greedy informed set within Rapidly-exploring Random Tree (RRT*)-like planners and analyze how greedy sampling affects exploration and asymptotic optimality. We then present Greedy RRT* (G-RRT*), a bi-directional anytime variant of RRT* that leverages the greedy informed set to focus sampling in the most promising regions of the search space. Experiments on abstract planning benchmarks, manipulation tasks from the MotionBenchMaker dataset, and a dual-arm Barrett WAM problem demonstrate that G-RRT* rapidly finds initial solutions and converges asymptotically to optimal paths, outperforming state-of-the-art sampling-based planners.

Phone Thiha Kyaw, Anh Vu Le, Rajesh Elara Mohan +1

Space & Terrestrial Autonomous Robotic Systems (STARS) Laboratory, University of Toronto Institute for Aerospace Studies, 4925 Dufferin Sreet, Toronto, M3H 5T6, Ontario, Canada. · Advanced Intelligent Technology Research Group, Faculty of Electrical and Electronics Engineering, Ton Duc Thang University, Ho Chi Minh City, 700000, Vietnam. · ROAR Lab, Engineering Product Development, Singapore University of Technology and Design, Singapore, 487372, Singapore.

Sampling-Based Motion Planning Tree Search

Ask PDF

Apr 8, 2026

Flow Motion Policy: Manipulator Motion Planning with Flow Matching Models

Open-loop end-to-end neural motion planners have recently been proposed to improve motion planning for robotic manipulators. These methods enable planning directly from sensor observations without relying on a privileged collision checker during motion planning. However, existing planners produce a single path for a given planning problem and cannot exploit their open-loop nature to propose multiple motion plans. To address this limitation, we introduce Flow Motion Policy, an open-loop neural motion planner that uses flow matching to generate a batch of motion plan proposals by learning a distribution over motion plans conditioned on the planning observation. At inference time, it samples multiple candidate motion plans to enable efficient best-of-

N

inference while avoiding iterative collision checking during planning. We benchmark the Flow Motion Policy against representative sampling-based, optimization-based and neural motion planning methods. Evaluation results demonstrate that Flow Motion Policy improves planning success and efficiency, highlighting the effectiveness of stochastic generative policies for end-to-end motion planning and best-of-

N

sampling. Project website: https://davoodsz.github.io/FlowMotionPolicy.github.io/

Davood Soleymanzadeh, Xiao Liang, Minghui Zheng

J. Mike Walker ’66 Department of Mechanical Engineering, Texas A&M University, College Station, TX 77843, USA · Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX 77843 USA

Robot Motion Planning Robot Motion Generation

Ask PDF

Abstract

Robot Motion Planning

Explore similar work

Performance Comparison of Classical and Neural Sampling Algorithms for Robotic Navigation

Greedy Heuristics for Sampling-Based Motion Planning in High-Dimensional State Spaces

Flow Motion Policy: Manipulator Motion Planning with Flow Matching Models