Friday 29 July 1994, 12:30pm, Wean 4601 Learning to Predict Interaction Forces for an Excavating Robot Sanjiv Singh I have a robot that digs in a sandbox. I give my planner a specification of a trench and it selects digs for the robot to perform. There are two interesting aspects to this task. First, since soil is diffuse, the state-space required to describe a goal is very high dimensional. Second, there are no good process (action) models for the task. That is, given an action, and a current state of the terrain, there are no models through which I could use to readily and accurately predict the effect of the action on the world. I will settle for something less. I would be happy to know the effect of the world on the robot. That is, if my planner could predict the resistive force experienced at the end effector of my digging robot, it could disqualify those digging trajectories that require forces beyond the capability of the robot. I have been looking into having my robot learn a model of the resitive force, from example. In the past few months I have examined a number of learning methods (GMBL, Projection Pursuit, Neural Nets). In this very informal talk, I will talk about how I have set up the learning problem and will discuss trade-offs among the methods that I have looked at.