Lockheed's robot is equipped with a 3D laser scanner that allows it to build detailed maps of its surroundings. It also has an array of acoustic sensors, which allow it to localize footsteps and voices. It can then combine the locations of humans with its 3D map to guess what areas the humans might be able to see, and then does its best to stay hidden. Keeping to the shadows, the robot always maintains an escape route, and if it senses a human approaching, it will look for the deepest darkest corner it can find and then hold its virtual breath until the danger has passed.
This is certainly not the first deceptive robot we've seen. Given the opportunity, a robot swarm at EPFL independently evolved the capacity for deception alarmingly quickly in a competition for virtual food. And researchers from Georgia Tech taught a robot to use deliberately deceptive tactics to fool other robots and humans. The Georgia Tech research, especially, seems like it's destined for applications like surveillance, as it endows a robot with a method of analyzing a situation to determine whether deception would be effective, based on what it knows about the robot (or person) trying to find it. If it decides that deception would help it achieve its goals, the robot then leaves tracks in one direction before moving off in a different one.
The key to avoiding detection by humans is to understand how you're perceived by humans. As Monty Python so astutely pointed out, for example, even the most perfect hiding place won't do you any good if it's the only possible place that you can be. By building models of both physical environments and perceptual environments, or how humans sense and react to things, robots will be much better at not just spying, but also understanding and reacting to us in less adversarial environments.