Robotik 1 - Kapitel 10 - Roboterprogrammierung und Programmierung durch Vormachen (Teil 2)
A robotics lecturer from HumanoidRobots explains how robots can learn skills and task sequences from human demonstrations, including how to capture, generalize, and transfer them to different robots.
In this talk
- 0:16
Learning by demonstration
The lecture introduces programming by demonstration as a way for robots to learn from people without requiring expert programming. The goal is to learn and generalize what a person demonstrates, rather than replay the same movements. Examples include wiping surfaces with different shapes and learning a sequence of actions to prepare dough or salad.
- 14:07
Challenges in imitation learning
The lecturer frames four key questions: who to imitate, when to imitate, what to imitate, and how to imitate. A robot may need to choose among teachers, identify the useful start and end of a demonstration, and decide whether the goal is to reproduce a movement or achieve its result. It must also map human actions to a robot with different sensors, body structure, and dynamics.
- 24:28
Capturing human movement
The lecture compares motion capture options, including marker based cameras, inertial sensors, exoskeletons, and markerless cameras. Vicon systems can track markers at high resolution but are expensive, require indoor setups, and can lose markers from view. Camera based methods need less specialized hardware but depend on complex tracking algorithms and have limited temporal resolution.
- 38:39
From trajectories to task models
Learning can happen at a trajectory level, where a robot generalizes a movement such as wiping, or at a symbolic level, where it learns a sequence of actions and their conditions. The lecturer describes segmenting demonstrations at key points, including changes in motion and changes in contact between hands and objects. For example, contact changes can help identify when a hand grasps, lifts, pours from, or places a bottle.
- 1:27:19
Transfer, feedback, and future directions
Learned skills must be mapped onto the robot and adapted using execution feedback, since human and robot kinematics differ. The lecturer highlights open research questions such as learning incrementally, learning from few or negative examples, and combining demonstrations with language, gaze, force, and audio. A cup touching a table, for instance, may be easier to detect through sound than vision alone.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.