Overview

Our motivation is to build autonomous robots that can follow people and recognize their activities. Such capabilities are important in applications such as home care for the elderly, intelligent environments, and iteractive media. In particular, the goal of this project is to develop reliable and efficient motion strategies for an autonomous robot to follow a target and keep it within the sensor range, despite occlusion by obstacles. We are currently investigating two approaches, POMDPs and greedy strategies.

Experimental setup: a Sick laser mounted on a Pioneer DX3 indoor robot.

POMDP Trackers

Target tracking has two variants that are often studied independently with different approaches: target searching requires a robot to find a target initially not visible, and target following requires a robot to maintain visibility on a target initially visible. We use partially observable Markov decision process (POMDP) to build a single model that unifies target searching and target following. The resulting POMDP policy exhibits interesting tracking behaviors, such as anticipatory moves that exploit target dynamics, information-gathering moves that reduce target position uncertainty, and energy-conserving actions that allow the target to get out of sight, but do not compromise tracking performance.

Some preliminary results are shown in the videos below. The problem setting is motivated by homecare applications. Imagine that an elderly person moves around at home and has a call button to call a robot over for help. The call status stays on for some time and then goes off. If the robot arrives while the call status is on, it gets a reward; otherwise, it gets no reward. Clearly the robot should stay close the person in order to improve the chance of receiving rewards, but at the same time, the robot needs to minimize movement in order to reduce power consumption. So the naive strategy of following right behind the person does not work well. What is interesting about these examples is that the robot manage to “track” the target while the target is outside the robot’s sensor visibility region most of the time.

Quicktime video (1.1 MB)	Quicktime video (1.0 MB)	Quicktime video (2.3 MB)
Light blue areas indicate obstacles. The green area around the robot indicate sensor’ visibility region. The various shades of gray show the robot’s belief of the target position. Lighter color indicates higher probability.

References

• D. Hsu, W.S. Lee, and N. Rong. A point-based POMDP planner for target tracking. In Proc. IEEE Int. Conf. on Robotics & Automation, pp. 2644–2650, 2008.
  BibTeX PDF

Greedy Followers

POMDP trackers integrate global information on the target behavior and the environment for optimal decision making. When little is known about the target behavior or the environment, a local greedy strategy is more effective. Key to our algorithm is the definition of a risk function, which tries to capture the target’s ability in escaping from the robot sensors’ visibility region in both short and long terms. To select actions effectively, the robot must balance between the short-term goal of preventing the immediate loss of the target and the long-term goal of keeping it visible for the maximum duration possible. Interestingly, a good comprise can be achieved, using only local information available to the robot’s sensors. By analyzing the local geometry, our algorithm computes a global risk function as a weighted sum of components, each associated with a single visibility constraint. It then chooses an action to minimize the risk locally in a greedy fashion.

As the algorithm uses only local geometric information available to the robot’s visual sensors, it does not require a global map and thus bypasses the difficulty of localization with respect to a global map. Furthermore, uncertainty in sensing and motion control does not accumulate. This improves the reliability of tracking.

This approach can be developed in both 2-D and 3-D. The 3-D case is, however, much more challenging technically, as both the robot and the target gain one more degree of freedom to maneuver and the visibility relationships in 3-D are more complex than those in 2-D. More details can be found here.

Quicktime video (0.6 MB)	Quicktime video (1.2 MB)
	Quicktime video (0.5 MB)	Quicktime video (0.5 MB)
The videos show the same motion of a robot helicopter from two different viewing angles.

References

• T. Bandyopadhyay, D. Hsu, and Ang Jr., M.H.. Motion strategies for people tracking in cluttered dynamic environments. In Proc. Int. Symp. on Experimental Robotics, 2008. To appear.
  PDF
• T. Bandyopadhyay, Ang Jr., M.H., and D. Hsu. Motion planning for 3-D target tracking among obstacles. In Proc. Int. Symp. on Robotics Research, 2007.
  PDF
• T. Bandyopadhyay, Y.P. Li, Ang Jr., M.H., and D. Hsu. A greedy strategy for tracking a locally predictable target among obstacles. In Proc. IEEE Int. Conf. on Robotics & Automation, pp. 2342–2347, 2006.
  PDF
• T. Bandyopadhyay, Y.P. Li, Ang Jr., M.H., and D. Hsu. Stealth tracking of an unpredictable target among obstacles. In M. Erdmann and others, editors, Algorithmic Foundations of Robotics VI—Proc. Workshop on the Algorithmic Foundations of Robotics (WAFR), pp. 43–58, Springer, 2004.
  PDF