SAUNA - Safe Autonomous Navigation

SAUNA is a major AASS 3-year project aimed at achieving international excellence in a research area of strong industrial relevance namely, safe autonomous navigation for professional industrial vehicles like forklift trucks, wheel loaders, mining trucks etc. SAUNA is Funded by the KK Foundation (www.kks.se) with a total of 10 MSEK for the period 2011-2014. SAUNA's research agenda has three specific objectives which are to be achieved in a close co-operation with a number of industrial partners.

Scientific objectives

1. Safe motion

To develop a model-based predictive controller to follow a predefined reference path/trajectory while moving at high speed (20-40 km per hour) and at the same time able to locally modify this trajectory to avoid collisions that may occur within a given preview horizon (short time horizon or collisions within seconds). The local modification of the trajectory should be such that the pre-determined final location can still be arrived at. Furthermore the predictive controller has to be able to account for:

a) Kinematic and dynamic constraints for a car-like vehicle, i.e., non-holonomic constraints, actuator limits, etc.

b) Constraints due to the work environment (could be both static and dynamic).

c) Task related constraints (may include temporal and resource constraints).

d) Limited computation time (and resources) for generating control inputs (related to high sampling rates).

2. Rich 3D perception

To develop a set of algorithms to create and maintain a single compact and rich 3D work-space model and use it for both safe navigation and planning. Rich 3D work-space model means 3D data points collected by a sensor(s) (e.g., laser scanner, time-of-flight camera etc.) and augmented with additional information such as color, semantic labels and confidence estimates. The model is compact:

- if it allows real-time operation of the vehicle, e.g., local obstacle avoidance within a given short-time horizon, localization in a given map within milliseconds, object detection in the milliseconds range;

- if it preserves all work-space features that can be used by all navigation-related algorithms e.g., mapping and localization (e.g., SLAM), static/moving objects detection, trajectory planning, collision prediction.

3. Hybrid planning

To develop a hybrid planner by integrating a symbolic task planner and a non-holonomic path/trajectory planner via the use of constraint-reasoning techniques. The hybrid planner has to obey a number of provable properties:

- it has to be correct: there exists a sequence of actions that lead to the task goal and this sequence satisfies certain time, resource and safety constraints; a very important constraint here is that the task plan can be executed within certain time interval;

- it has to be optimal: a task plan may be optimal from the viewpoint of the task itself, e.g., small number of sub-tasks; it also can be optimal with respect to each path/trajectory, e.g., shorter path for each sub-task; however such a task plan may not be globally optimal, i.e., there may be another task plan that has more sub-tasks, but the total path length is shorter;

- it has to be executable, i.e., the pats/trajectories planned must obey the geometric, kinematic and dynamic aspects of motion for non-holonomic vehicles.