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

TL;DR

This paper introduces a novel "human-like planning" (HLP) algorithm that learns to navigate cluttered environments by observing human demonstrations in VR. By abstracting human strategies into qualitative task representations, HLP generates high-level plans that are more efficient and scalable than traditional random sampling methods, outperforming a state-of-the-art planner in simulations and real-world experiments.

What Problem Does This Paper Address?

Robots struggle with planning efficient paths in cluttered environments, especially when many objects are involved. Traditional methods like sampling-based planners become computationally expensive as the dimensionality of the problem increases with the number of objects. Humans, on the other hand, are remarkably adept at these tasks, often finding intuitive and quick solutions. This paper aims to bridge this gap by learning from human expertise to create robot planners that are more efficient and human-like in their decision-making.

Methods

The core of the HLP approach involves learning from human demonstrations captured in a virtual reality (VR) environment. The process can be broken down into these key steps:

  • Human Demonstration Capture: Participants performed reaching tasks in VR, with their arm movements and the environment state recorded.
  • Qualitative Task Space Modeling: The environment is abstracted into a qualitative representation, focusing on "occupied space," "gaps" (free space between objects), and "connecting space." This abstraction is crucial for generalizing beyond specific object configurations.
  • Action Space Discretization: Human actions are categorized into two primitive actions: "moving an object" and "going through a gap."
  • Demonstration Segmentation and Feature Extraction: Human trajectories are segmented into state-action pairs. Features are extracted based on spatial relationships like distances, sizes, and orientations between objects, gaps, and the target.
  • Decision Classifier Training: These features are used to train binary classifiers (e.g., for selecting a gap or an object, and for determining object movement direction).
  • High-Level Plan Generation: The trained classifiers are used by the HLP algorithm to generate a sequence of waypoints representing a high-level plan in the task space.
  • Robot Model Transfer: This high-level plan can then be transferred to any robot model and used to initialize a low-level controller for inverse kinematics and trajectory optimization.

Datasets

  • Human VR Dataset: A dataset was collected from 24 participants performing 90 reaching trials each in a cluttered VR environment. This dataset was split into training (80%) and testing (20%) sets, with a cross-participant split to ensure generalization.
  • Robot Simulation Data: Experiments were conducted in a physics-based robot simulation environment.
  • Real Robot Data: Experiments were performed using a UR5 arm on an omnidirectional robot.

Key Insights and Ideas

  • Human Expertise as a Blueprint: The paper highlights that humans possess intuitive strategies for navigating clutter that are not easily captured by purely geometric or random sampling methods. Learning these high-level strategies is key.
  • Qualitative Abstraction for Scalability: By moving from a high-dimensional configuration space to a qualitative task space representation (rows, gaps, connecting space), the planner becomes less sensitive to the exact number and configuration of obstacles. This is a significant step towards scalable planning.
  • Bridging High-Level and Low-Level Control: The HLP generates a high-level plan that acts as a guide for a low-level controller, effectively combining abstract reasoning with concrete execution. This modularity allows it to work with various robot platforms.
  • Generalization Beyond Training Data: The approach demonstrates strong generalization capabilities, performing well on unseen VR data, diverse simulation scenarios, and real-world robot tasks, even with different object counts and arrangements.

Key Results

  • VR Data Similarity: The HLP achieved an average similarity of 70% to specific human participant plans in unseen VR trials. It selected the same action 79% of the time and the same element (gap or object) 67% of the time.
  • Robot Simulation Performance: In simulation, the HLP significantly outperformed a standard physics-based stochastic trajectory optimization (STO) approach. For example, in one generalisation test, HLP achieved a 94% success rate with a total time of 1.56 seconds, compared to STO's 84% success rate and 17.88 seconds total time.
  • Planning Time Efficiency: A key advantage of HLP is its consistent, low "initialization time" for planning, which remains nearly fixed regardless of the number of obstacles. This contrasts with STO, where planning time increases significantly with complexity.
  • Real Robot Success: Real robot experiments showed HLP favoring free space navigation, while STO was often forced to push objects due to its straight-line initialization bias.

Limitations and Future Work

  • Limitations: The paper acknowledges that the current approach is primarily focused on reaching tasks. The qualitative representation of rows is a simplification that might not apply to all cluttered environments.
  • Future Work: The authors suggest several avenues for future research, including:
  • Augmenting training data with more generalized scenes.
  • Using more powerful classifiers.
  • Developing closed-loop planning.
  • Investigating a more detailed mechanistic understanding of human planning and execution.
  • Extending the approach to more complex manipulation tasks beyond simple reaching.

Summary updated

Abstract

Humans, in comparison to robots, are remarkably adept at reaching for objects in cluttered environments. The best existing robot planners are based on random sampling of configuration space- which becomes excessively high-dimensional with large number of objects. Consequently, most planners often fail to efficiently find object manipulation plans in such environments. We addressed this problem by identifying high-level manipulation plans in humans, and transferring these skills to robot planners. We used virtual reality to capture human participants reaching for a target object on a tabletop cluttered with obstacles. From this, we devised a qualitative representation of the task space to abstract the decision making, irrespective of the number of obstacles. Based on this representation, human demonstrations were segmented and used to train decision classifiers. Using these classifiers, our planner produced a list of waypoints in task space. These waypoints provided a high-level plan, which could be transferred to an arbitrary robot model and used to initialise a local trajectory optimiser. We evaluated this approach through testing on unseen human VR data, a physics-based robot simulation, and a real robot (dataset and code are publicly available <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> ). We found that the human-like planner outperformed a state-of-the-art standard trajectory optimisation algorithm, and was able to generate effective strategies for rapid planning- irrespective of the number of obstacles in the environment.
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