Probabilistic Mobile Manipulation in Dynamic Environments, with Application to Opening Doors

Anna Petrovskaya, Andrew Y. NgIJCAI 2007

Introduction

Mobile manipulation—the ability for a robot to interact with objects while moving—remains a significant challenge in robotics. Traditionally, navigation and manipulation have been treated as separate domains. Navigation typically relies on coarse occupancy grids (e.g., 10cm resolution) sufficient for avoiding walls, whereas manipulation requires sub-centimeter precision (1–5mm) to operate handles or grasp objects. Furthermore, standard Simultaneous Localization and Mapping (SLAM) algorithms often assume a static environment. When a robot opens a door, it alters the map, which can lead to localization failure.

This paper introduces a unified probabilistic framework that allows a mobile robot to simultaneously localize itself and estimate the state of objects it interacts with, such as the angle of a door. By bridging the gap between coarse navigation and fine-grained manipulation, the system enables robots to perform complex tasks in dynamic office environments.

Unified Probabilistic Framework

The core of the proposed approach is a Dynamic Bayesian Network (DBN) that tracks both the robot’s pose and the object’s state (e.g., the door angle $\alpha$).

Dynamic Bayesian Network for tracking robot pose $x_t$ and object state $\alpha_t$ given controls $u_t$ and observations $z_t$.

To manage the high-dimensional state space, the researchers employed a Rao-Blackwellized Particle Filter (RBPF). In this setup, each particle represents a potential robot trajectory. Attached to each particle is an Extended Kalman Filter (EKF) that maintains the estimate of the door's state. This factorization allows the robot to treat manipulation as a localized SLAM problem: as the door moves, the robot updates its "dynamic map" rather than becoming lost.

Multi-Resolution Representation and Model Pasting

A key innovation is the use of a multi-resolution map. Instead of attempting to map an entire building at 1mm resolution—which would be computationally prohibitive—the system uses a standard 10cm occupancy grid for general navigation. High-resolution (0.1cm) polygon models of specific objects, like doors, are then "pasted" onto this coarse map.

Multi-resolution approach: (a) Coarse global map, (b) STAIR robot in a hallway, (c) High-resolution object model integrated into the navigation map.

This "model pasting" leverages the repetitive nature of office environments. A single high-fidelity model of a door can be reused across the entire map. To ensure sensor data remains consistent across these varying resolutions, the authors implemented a probabilistic "opacity" model, ensuring that laser ray blockage probabilities are uniform regardless of cell size.

Achieving Sub-Centimeter Precision

Standard particle filters often struggle with highly "peaked" distributions where the correct state is a very precise point. To overcome this, the authors introduced the Scaling Series (SS) algorithm. This technique starts with high simulated sensor noise and gradually "anneals" or reduces it. This process guides the particles toward the high-probability region, allowing the robot to achieve sub-centimeter precision with as few as 100 particles.

Experimental Results

The algorithm was tested on the STAIR (STanford Artificial Intelligence Robot) platform across 12 distinct scenarios involving four different doors. The results demonstrated a significant leap in accuracy:

  • Standard AMCL: Failed significantly when the door state didn't match the map, showing an overall RMS error of 42.975 cm.
  • Standard RBPF: Performed better but still lacked the precision for manipulation, with an error of 1.779 cm.
  • Proposed RBPF-SS: Achieved a remarkable RMS error of 0.358 cm (approx. 3.5 mm).

Qualitative tests showed the STAIR robot successfully navigating hallways, identifying doors, grasping handles, and entering rooms while maintaining precise localization throughout the manipulation process.

Conclusion and Future Work

By treating manipulation as an extension of navigation, this framework allows robots to operate effectively in semi-static environments. Future work will focus on extending this model to more diverse objects, such as filing cabinets or movable furniture, and improving robustness in highly crowded environments where unmodeled moving objects (like people) are prevalent.

Summary updated

BibTeX
@inproceedings{petrovskaya2007probabilistic,
  author = {Anna Petrovskaya and Andrew Y. Ng},
  title = {{Probabilistic Mobile Manipulation in Dynamic Environments, with Application to Opening Doors}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2007},
}
Figure