Stanford Seminar - Deploying Autonomous Service Mobile Robots, And Keeping Them Autonomous
A Stanford Online seminar presenter from the University of Texas at Austin discusses how to deploy service robots in changing human environments and help them remain reliable over time.
In this talk
- 0:11
A vision for everyday service robots
The goal is to let non-specialists deploy affordable, general-purpose mobile robots in places such as campuses, businesses, and homes. These robots should handle useful tasks, from checking trees that need attention to delivering packages, and keep working reliably over years despite unfamiliar conditions.
- 5:02
Localizing amid a changing world
Tables, chairs, construction barriers, and parked cars can make most sensor observations differ from a robot’s static map. Episodic non-Markov localization separates stable map features from temporary and changing ones, adjusting how it uses observations as conditions shift. Probabilistic Object Maps add learned spatial patterns, such as where cars tend to park, to help localization when objects move.
- 21:48
Recognizing when perception may fail
Robots can make risky choices because their algorithms do not account for when their own inputs may be misleading. Introspective perception learns to flag unreliable image regions, including reflections, shadows, glare, and moving people; in the visual SLAM example, IB-SLAM filtered or downweighted troublesome features so the robot could keep tracking its position.
- 39:11
Learning routes and behaviors from people
For path planning, human driving examples teach a robot which terrain people prefer without requiring labeled terrain classes. For decisions such as waiting for a door or navigating social situations, physics-informed program synthesis turns demonstrations into policies based on meaningful quantities such as position and velocity. The presenter reports that a soccer policy reached 87% accuracy and could be improved with a small number of corrections.
- 55:52
Next steps for long-term autonomy
Future work includes questioning whether robots need metric maps, helping them test their own limits, and combining awareness of competence with task performance. The presenter also points toward reasoning about semantic entities, since objects such as chairs remain recognizable through lighting and viewpoint changes that can disrupt low-level visual features.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.