Dexter Learning Demos

Learning to Reach: From Manual Skill to Communicative Action

The following is a demonstration that it is possible for a robot configured to learn manual skills, with no direct support for communication, to learn the utility of an important communicative action – pointing. Ths achieved through the several stages:

  1. A ball is placed within the reachable workspace of the robot ALWAYS. Through exploration the robot discovers the appropriate sequence of primitive actions that leads to the rewarding tactile sensation. This process is analogus to the “motor babbling” stage during infant development.
  2. During this stage, objects are placed sometimes out-of-reach of the robot. This causes a noticable change in the previously learned policy’s transition dynamics. The robot autonomous adapts to this change by discovering the position of the object with respect to the success of a the “reach” action, thus learning a new policy (not to reach for out-of-reach objects) to increase its rate of reward.
  3. The dynamics of the environment is changed yet again during this stage – a human enters the scene and causes the out-of-reach object sometimes appear within reach of the robot when an indication of assistance request is observed through the robot’s actions. Again, the robot notice the change in the environmental dynamics and discovers the appropriate cues and corresponding optimal action sequence such that rate of reward can increase once again.

embedded by Embedded Video

YouTube link to: Stage 1: Learns to reach
embedded by Embedded Video

YouTube link to: Stage 2: Learns to differentiate reachable and non-reachable workspace
embedded by Embedded Video

YouTube link to: Learns to point to request for assistance in the presence of human-scale “objects”

Compared with the programming interface. this approach is a natural and intuitive way for programming the robot – through manipulation of environment and acting through common sense to induce the emergence of appropriate actions, including ones that are communicative: making this negotiation process intuitive enough such potentially even uninformed human subjects can go through.

Other Dexter Demonstrations
joint work with John Sweeney and Stephen Hart

  1. A simple task is demonstrated to Dexter through teleoperation. The robot infers intentions of the human operator by observing his actions. The task is replayed by Dexter in the end.
  2. The sorting task taught via teleoperation is replayed with previously learned implicit common sense knowledge incorporated. Dexter determines handedness intelligently given the position of the object and the target location, without being explicitly taught by the operator. Thus, saving much energe for the teleoperator.

embedded by Embedded Video

YouTube link to: First teach Dexter sorting via teleoperation, and then replays task with previously learned common sense knowledge
embedded by Embedded Video

YouTube link to: Dexter handling another sorting task using acquired common sense knowledge

Fun Stuffs

Related work

Leave a Reply