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

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.

Highlights

Inject planning structure directly into transformer attention.

GAIDE turns robot embodiment and workspace geometry into a graph, then uses that graph as an attention mask so the neural sampler reasons over structurally meaningful robot-scene interactions.

Graph Masking

encode kinematic and spatial structure as attention constraints inside the transformer.

4.81

lowest average planning cost across all evaluated planners and tasks.

58%

average success rate across six held-out planning environments.

Zero

additional fine-tuning needed for real-world UR5e deployment.

Method

Graph attention makes neural sampling spatial- and embodiment-aware.

At each planning step, GAIDE observes the current configuration, goal configuration, robot point cloud, and workspace point cloud, then predicts a delta joint angle to extend the bidirectional tree.

A → masked attention → δq_t

1

Construct the Graph

Build embodiment edges along the robot and spatial edges from workspace nodes to robot nodes.

2

Mask Attention

Use the graph adjacency matrix to restrict transformer attention to meaningful token pairs.

3

Grow the Planner

Decode a goal-directed joint update and use it as an informed sample inside the planner.

Graph construction over robot and workspace point clouds

Graph Construction. Robot nodes follow the manipulator topology, while workspace nodes connect to robot nodes to encode spatial obstacle relationships.

GAIDE Architecture. Point-cloud and configuration tokens are fused by a transformer encoder with interleaved masked and vanilla attention layers before decoding the next joint update.

Results

We evaluate GAIDE on six held-out planning tasks against three classes of baselines: uniform samplers (Bi-RRT, RRT*), heuristic-based informed samplers (IRRT*, BIT*), and neural informed samplers (MPNets, SIMPNet). All planners share the same bidirectional planning framework and PyBullet collision checker. Classical planners are given a time budget equal to GAIDE's average planning time per task. Three metrics are reported: planning time T [s], success rate S [%], and planning cost C (path length in configuration space).

Evaluation Environments

Figure 3 — Held-out Planning Tasks. Six environments spanning a range of geometric complexity: TableTop, Box, and Bins (top row); Shelf I, Shelf II, and Shelf III (bottom row, increasing number of shelves).

Planning Metrics

Interactive comparison of success rate S [%], planning time T [s], planning cost C, and average success/time across all held-out tasks. Toggle the metric to compare uniform, heuristic, and neural informed samplers.

Ablation Study

We ablate the attention masking design with two controlled variants to isolate the contribution of the interleaved masking strategy:

GAIDE-V (Vanilla Transformer). Standard encoder-decoder transformer with no attention masking — processes all tokens uniformly without encoding spatial or kinematic structure.
GAIDE-H (Hard masking). Applies the graph attention mask at every encoder layer rather than interleaving. This prevents the decoder from fully attending to workspace embeddings, causing spatial information to be masked out at each layer.

GAIDE's interleaved design — alternating masked and vanilla layers — consistently outperforms both variants, confirming that the balance between structured and global attention is critical.

Interactive ablation of the attention masking strategy. GAIDE interleaves masked and vanilla attention layers, while GAIDE-V removes masking and GAIDE-H applies hard masking at every layer.

Real-World Deployment

We deploy GAIDE on a physical UR5e robotic arm without any additional training or fine-tuning. The scene is represented using point cloud data from a calibrated Intel RealSense D435i RGB-D camera, with the camera-to-robot-base transform estimated via AprilTag markers. Collision checking uses a spherical approximation of the robot geometry. Three demonstrations are shown: a bin task (Demo I) and two shelf-insertion tasks of increasing difficulty (Demo II, Demo III). Each demo shows the physical robot (top) alongside the corresponding simulated point cloud viewed by the planner (bottom).

Figure 6 — Real-World Deployment. Demo I: bin task. Demo II & III: shelf-insertion tasks with increasing obstacle complexity. Top rows show the physical UR5e robot; bottom rows show the simulated point cloud scene used by the planner. GAIDE plans collision-free paths in all three scenarios without retraining.

BibTeX

@article{soleymanzadeh2026gaide,
  title   = {GAIDE: Graph-based Attention Masking for Spatial- and Embodiment-aware Motion Planning},
  author  = {Soleymanzadeh, Davood and Liang, Xiao and Zheng, Minghui},
  journal = {arXiv preprint arXiv:2603.04463},
  year    = {2026}
}