Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DRIFT: Diffusion-based Rule-Inferred For Trajectories

Jinyang Zhao, Handong Zheng, Yanjiu Zhong, Qiang Zhang, Yu Kang, Shunyu Wu

TL;DR

This article proposes DRIFT (Diffusion-based Rule-Inferred for Trajectories), a conditional diffusion framework designed to generate high-fidelity reference trajectories by integrating two complementary inductive biases, achieving centimeter-level imitation fidelity and competitive smoothness.

Abstract

Trajectory generation for mobile robots in unstructured environments faces a critical dilemma: balancing kinematic smoothness for safe execution with terminal precision for fine-grained tasks. Existing generative planners often struggle with this trade-off, yielding either smooth but imprecise paths or geometrically accurate but erratic motions. To address the aforementioned shortcomings, this article proposes DRIFT (Diffusion-based Rule-Inferred for Trajectories), a conditional diffusion framework designed to generate high-fidelity reference trajectories by integrating two complementary inductive biases. First, a Relational Inductive Bias, realized via a GNN-based Structured Scene Perception (SSP) module, encodes global topological constraints to ensure holistic smoothness. Second, a Temporal Attention Bias, implemented through a novel Graph-Conditioned Time-Aware GRU (GTGRU), dynamically attends to sparse obstacles and targets for precise local maneuvering. In the end, quantitative results demonstrate that DRIFT reconciles these conflicting objectives, achieving centimeter-level imitation fidelity (0.041m FDE) and competitive smoothness (27.19 Jerk). This balance yields highly executable reference plans for downstream control.

DRIFT: Diffusion-based Rule-Inferred For Trajectories

TL;DR

This article proposes DRIFT (Diffusion-based Rule-Inferred for Trajectories), a conditional diffusion framework designed to generate high-fidelity reference trajectories by integrating two complementary inductive biases, achieving centimeter-level imitation fidelity and competitive smoothness.

Abstract

Trajectory generation for mobile robots in unstructured environments faces a critical dilemma: balancing kinematic smoothness for safe execution with terminal precision for fine-grained tasks. Existing generative planners often struggle with this trade-off, yielding either smooth but imprecise paths or geometrically accurate but erratic motions. To address the aforementioned shortcomings, this article proposes DRIFT (Diffusion-based Rule-Inferred for Trajectories), a conditional diffusion framework designed to generate high-fidelity reference trajectories by integrating two complementary inductive biases. First, a Relational Inductive Bias, realized via a GNN-based Structured Scene Perception (SSP) module, encodes global topological constraints to ensure holistic smoothness. Second, a Temporal Attention Bias, implemented through a novel Graph-Conditioned Time-Aware GRU (GTGRU), dynamically attends to sparse obstacles and targets for precise local maneuvering. In the end, quantitative results demonstrate that DRIFT reconciles these conflicting objectives, achieving centimeter-level imitation fidelity (0.041m FDE) and competitive smoothness (27.19 Jerk). This balance yields highly executable reference plans for downstream control.
Paper Structure (29 sections, 8 equations, 5 figures, 1 table, 1 algorithm)

This paper contains 29 sections, 8 equations, 5 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Prior Work and Background
  3. Structured Scene Encoding for Trajectory Generation
  4. Generative Models for Motion Planning
  5. Optimizing Smoothness and Precision in Motion Planning
  6. PROBLEM FORMULATION
  7. METHODOLOGY
  8. Overview
  9. Relational Inductive Bias: Structured Scene Perception
  10. Temporal Inductive Bias: Recurrent Denoising Generator
  11. Dynamic Obstacle Attention
  12. Target-Driven Gating
  13. Training Objectives and Curriculum
  14. Imitation Fidelity ($\mathcal{L}_{imit}$)
  15. Kinematic Coherence ($\mathcal{L}_{smooth}$)
  16. ...and 14 more sections

Figures (5)

  • Figure 1: DRIFT fuses relational topological bias (SSP) and temporal attention bias (GTGRU) to produce high-fidelity reference plans. The Denoising Network (DN) then iteratively refines trajectories from these structured encodings, balancing smoothness and precision.
  • Figure 2: Voxel downsampling converts the dense LiDAR point cloud (green) into sparse graph nodes (red) (Alg. \ref{['alg:ssp_module']}), cutting computational cost and producing structured topologies for GNN relational reasoning.
  • Figure 3: Terminal accuracy comparison for close-proximity docking. Our method, DRIFT (purple), maintains 92.55% success at 0.1m, while the DTG baseline (blue) fails (45.49%). This result demonstrates DRIFT's ability to resolve the "precision-smoothness trade-off" by achieving high-precision local maneuvering.
  • Figure 4: Predicted trajectories (dark) vs. ground truth (light gray) across models: DTG is smooth but deviates and misses terminals; CAVE and BC are erratic or imprecise. DRIFT resolves the precision–smoothness trade-off, matching the ground truth while remaining smooth.
  • Figure 5: Predicted trajectory (dark) closely matches ground truth (light gray) in (a) trajectory, (b) 2D map, and (c) point-cloud views, showing DRIFT reconciles precision and smoothness.