Learning from Demonstration

Videos of learned motions

Description

In the space environment, teleoperating robots like NASA's Robonaut can often be slow and tedious due to communications time delay. Instead, I envision a robot that can recognize a teleoperator's intended motion and autonomously continue the execution of recognized routine tasks, for either the remainder of the task, or enough to reduce an operator's perceived time delay. To do this, the robot learns offline a library of generalized activities from a training set of user demonstrations. During online operations, the robot can potentially perform real-time recognition of a user's teleoperated motions, and if requested, autonomously execute the remainder of an activity.

My doctoral thesis presents an approach to learning complex physical motions from human demonstration that (1) provides flexibility during execution while robustly encoding a human's intended motions, and (2) automatically determines the relevant features of a motion so that they can be preserved during autonomous execution in new situations.

I also introduce an approach to real-time motion recognition that (1) leverages temporal information to successfully model motions that may be non-Markovian, (2) provides fast real-time recognition of motions in progress by using an incremental dynamic time warping approach, and (3) employs the probabilistic flow tube representation that enables our method to recognize learned motions despite varying environment states.

Human Intent Recognition

My MS thesis enables a new kind of interaction between humans and computational agents, such as robots or computers, by allowing the agent to anticipate and adapt to human intent. In the future, more robots may be deployed in situations that require collaboration with humans, such as scientific exploration, search and rescue, hospital assistance, and even domestic care. These situations require robots to work together with humans, as part of a team, rather than as a stand-alone tool. The intent recognition capability is necessary for computational agents to play a more collaborative role in human-robot interactions, moving beyond the standard master-slave relationship of humans and computers today.

I develop an innovative capability for recognizing human intent, through statistical plan learning and online recognition. I approach the plan learning problem by employing unsupervised learning to automatically determine the activities in a plan based on training data. The plan activities are described by a mixture of multivariate probability densities. The number of distributions in the mixture used to describe the data is assumed to be given. The training data trajectories are fed again through the activities' density distributions to determine each possible sequence of activities that make up a plan. These activity sequences are then summarized with temporal information in a temporal plan network, which consists of a network of all possible plans. The approach to plan recognition begins with formulating the temporal plan network as a hidden Markov model. Next, the most likely path can be determined by using the Viterbi algorithm. Finally, I refer back to the temporal plan network to obtain predicted future activities.

This research presents several innovations: First, I introduce a modified representation of temporal plan networks that incorporates probabilistic information into the state space and temporal representations. Second, I learn plans from actual data, such that the notion of an activity is not arbitrarily or manually defined, but is determined by the characteristics of the data. Third, I develop a recognition algorithm that can perform recognition continuously by making probabilistic updates. Finally, the recognizer not only identifies previously executed activities, but also predicts future activities based on the plan network.

I demonstrate the capabilities of the algorithms on motion capture data. The results show that the plan learning algorithm is able to generate reasonable temporal plan networks, depending on the dimensions of the training data and the recognition resolution used. The plan recognition algorithm is also successful in recognizing the correct activity sequences in the temporal plan network corresponding to the observed test data.

For more details, please see my MS thesis (pdf).

Research Group

I work in the Computer Science and Artificial Intelligence Lab (CSAIL), with the Model-based Embedded and Robotic Systems (MERS) group. We work with a variety of hardware platforms including the Barrett Whole Arm Manipulator robot, the Willow Garage PR2 humanoid robot, and quadrotor helicopters. Our group is headed by Professor Brian Williams.