Human-like Planning for Reaching in Cluttered Environments

Mohamed Hasan, Matthew Warburton, Wisdom C. Agboh, Mehmet Remzi Dogar, Matteo Leonetti, He Wang, Faisal Mushtaq, Mark Mon-Williams, Anthony G. Cohn

Abstract

The Human-like Planner (HLP) bridges the gap between human intuition and robotic precision by learning high-level strategies from human demonstrations in Virtual Reality. By abstracting cluttered environments into qualitative spatial representations, the HLP generates effective manipulation plans that outperform state-of-the-art trajectory optimizers in both success rate and planning speed, regardless of the number of obstacles.

Introduction

Robotic navigation in cluttered environments, such as reaching into a crowded refrigerator, often suffers from the "curse of dimensionality." Traditional Sampling-Based Planners (SBPs) struggle with computational costs as obstacle density increases. In contrast, humans exhibit an intuitive ability to decide whether to navigate through gaps or manipulate obstacles. This paper introduces the Human-Like Planner (HLP), a framework designed to capture this qualitative logic and transfer it to robotic motion planning.

Methods

The HLP framework utilizes a multi-stage approach to integrate human strategic logic with robotic execution.

Overview of the Human-Like Planner (HLP) framework, showing the pipeline from VR data collection to real-world execution.

The process begins with VR Data Capture, recording 2,160 reaching trials to capture natural human strategies. The environment is then modeled using Qualitative Spatial Representations (QSRs), partitioning the workspace into rows, gaps, and objects.

Qualitative Spatial Representation (QSR) partitioning the workspace into gaps, segments, and objects to facilitate human-like decision making.

Decision Classifiers (SVMs and GPR) are trained on segmented human data to predict high-level actions, such as which gap to traverse or which object to move. Finally, a Hierarchical Planning stage generates a task-space "skeleton" plan, which a low-level controller translates into robot-specific trajectories.

Key Insights

  • Abstraction and Generalization: By focusing on qualitative relations (e.g., "between object A and B") rather than precise coordinates, the planner becomes object-agnostic. This allows it to generalize to scenes with varying numbers of obstacles without a significant increase in planning time.
  • Strategic Initialization: A major hurdle for trajectory optimizers is a poor starting guess. The HLP provides a "human-inspired" initial path that is already close to a feasible solution, significantly accelerating convergence.
  • Logic Transfer: The system learns the logic of the decision-making process rather than just cloning trajectories. This makes the high-level plan transferable to any robot model, regardless of its specific joint configuration.

Results

Evaluation across simulation and real-world UR5 setups demonstrated that the HLP achieves a 94% success rate, significantly higher than the 84% achieved by state-of-the-art Stochastic Trajectory Optimizers. Furthermore, the HLP reduced planning time from 17.88s to 1.56s in structured scenes. Notably, the HLP's planning time remains nearly constant as clutter increases, demonstrating superior scalability over traditional methods.

Conclusion

The HLP demonstrates that qualitative abstraction of human decision-making can significantly enhance robotic planning in clutter. Future work will focus on extending the framework to unstructured 3D environments and implementing closed-loop control for real-time adaptation.

Summary updated

BibTeX
@inproceedings{hasan2020human,
  author = {Mohamed Hasan and Matthew Warburton and Wisdom C. Agboh and Mehmet Remzi Dogar and Matteo Leonetti and He Wang and Faisal Mushtaq and Mark Mon-Williams and Anthony G. Cohn},
  title = {{Human-like Planning for Reaching in Cluttered Environments}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2020},
  url = {https://openreview.net/forum?id=76pcdIuNSeT},
}
Figure