Planning and Perception for Robotic Manipulators in Semi-structured Environments
- Publication Type:
- Thesis
- Issue Date:
- 2022
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
New applications of robotic manipulators require a far greater degree of autonomy than traditionally has been needed. Industrial robot manipulators such as collaborative robots are designed for advanced manufacturing applications where they should operate safely in dynamic work environments shared with humans and should be able to adapt quickly to perform a variety of tasks. Manipulators for fruit harvesting in agricultural robotics need to operate quickly in semi-structured outdoor environments in order to perform at a commercially viable level.
These applications require many components to work together robustly and efficiently in order to be viable for real world use. For example, robotic harvesting involves task and motion planning, task space control, active perception, object segmentation and grasping. Extending this to multiple robots requires coordination among agents which is non-trivial and further increases problem complexity. Existing methods for solving such problems in real-time often resort to making coarse approximations to parts of the problem, for example a simplified motion or sensor model. This can lead to undesirable behaviour particularly for highly non-linear systems such as robotic arms or complex sensor models such as those for RGB-D cameras. Furthermore, when searching for objects it remains a challenge to effectively balance between exploration of a scene and targeted observation of key regions.
This thesis is concerned with developing component algorithms that solve such problems effectively and exhibit computational behaviour that is well-understood analytically. Key to achieving this is the exploitation of known a priori structure in the problem, typical of semi-structured environments. First, a novel planning framework called Hausdorff approximation planner (HAP) is described which utilises prior knowledge for building a library of trajectories which are guaranteed to be of bounded length and approximately correspond to geodesics in the task space. These trajectories can be concatenated reliably over long task horizons, adapted quickly online and queried for accurate costs in constant time. Then, a novel targeted information gathering algorithm is presented which is integrated with HAP and other components into a non-myopic active perception framework suitable for systems with complex motion and sensor models and effectively balances between exploration and targeted observations of the scene. This is then generalised into a decentralised coordination framework suitable for flexible multi-robot perception problems that facilitate downstream processes, such as manipulation, object segmentation, and collaborative tasks. Demonstration experiments include robotic pruning, robotic harvesting, and searching for objects in clutter using multiple robotic manipulators with eye-in-hand sensors.
Please use this identifier to cite or link to this item: