A cute little robot happily drives around a box avoiding obstacles until it starts to slow down as its batteries run low. Realizing this, the robot turns towards a light over its charging station and goes in to "feed." By looking at its behavior, we might be led to believe that the robot was finding out about obstacles and avoiding them until it got "hungry." However, this vehicle has no understanding or representation of the world it lives in. Instead, it is reacting to elements of its environment such as bump sensors, directional light level, and its own battery voltage. In the early 1950's, W. Grey Walter [1] built a number of robots (which he called tortoises) which operated in this fashion. Amazingly, their complex behavior derived from only a small number of relays, vacuum tubes, sensors, actuators, and, most importantly, their interaction with their environment.
Like Walter's tortoises, a situated robot is one which does not deal with abstract representations of the world (which may be simulated or real), but rather reacts directly to its environment as seen through its sensors [2]. An alternative to having a situated robot would be one which builds up a representation of its world and then makes plans based on the representation. Because of the limitations of our present technology, these two approaches often seem contradictory. In the present, each approach is better for different applications. If we want to make an "artificial person" at some point in the future, we will need to incorporate both approaches.
The situated approach can only deal with a small domain of problems. When a robot gets into a situation where it needs to reason or plan ahead in order to reach a goal, simply reacting to its environment is insufficient. A situated robot can be thought of as sensors and goals feeding into a fixed network of difference engines which have actuators as their outputs. For example, we might have a light sensors and a goal of reaching lights fed into a difference engine. The robot would then activate its actuators in an attempt to reach its goal. This network might be slightly more complex with a controller suppressing and activating difference engines in response to an FSM state or a sensor input (for example, causing the robot to flee when danger is detected). However, this architecture does not give us the flexibility we need to solve more complicated problems, such as figuring out that we need to move further from some goals in order to reach an overall goal. A situated robot might have special cases built into it (e.g. for dealing with getting stuck in a corner or down a dead-end hallway), but it would be very difficult to deal with the general case this way.
If the situated approach is so limited, why should anyone spend any effort on it? From a purely practical perspective, it can be used to deal with small problem domains without massive amounts of processing power. If all we need is a robot that can drive around a building while avoiding walls (carrying messages around, for example) then the situated approach may work just fine. High level machine vision is still a hard problem, making it difficult to build an accurate representation of the world. Situated robots can also be made with small, low cost microprocessors. It's also likely that most lower life forms don't have a good representation of the world around them, but instead operate using situated mechanisms.
Higher life forms, such as humans, build up representations of the world around themselves and make plans and act based upon them. However, even in higher life forms many lower level agencies still operate using a situated approach. For many applications, the representation approach can't close control loops fast enough. For example, when we move our head rapidly, our eyes are kept looking in the same direction by sensors in our inner ears. Building up a representation of the world through vision and using it to keep our eyes stabilized would be too slow. As another example, we pull our hand away when we touch a hot object without doing any reasoning.
The situated approach is good for dealing with problems where planning ahead is unnecessary or takes too much time. However, the representation approach is needed for solving more complicated problems where it is necessary to reason about the state of the world. Research into both fields is valuable. For dealing with complicated tasks in the real world, it will probably be necessary to fuse the two approaches. Reasoning can be used to build up higher level plans and solve high level problems while lower level agencies may use a more situated approach for carrying out plans and dealing with problems which need immediate attention.
[1]
W. G. Walter, The Living Brain.
Duckworth, London, United Kingdom. 1953.
[2]
R. A. Brooks and C. Ferrell, Embodied Intelligence.
MIT Press, Cambridge, MA. DRAFT 1995.
Erik Nygren
(nygren@mit.edu)