Towards Continual Topological Mapping with Introspection

Speaker: Rohan Paul , Oxford University

Date: Thursday, April 10, 2014

Time: 10:00 AM to 11:00 AM Note: all times are in the Eastern Time Zone

Refreshments: 9:45 AM

Public: Yes

Location: 32-D463 (Star)

Event Type:

Room Description:

Host:

Contact: Nick Roy, nickroy@csail.mit.edu

Relevant URL:

Speaker URL: None

Speaker Photo:
None

Reminders to: robotics-seminars@csail.mit.edu

Reminder Subject: TALK: Towards Continual Topological Mapping with Introspection

Abstract:

For robust, life-long autonomous operation in dynamic unstructured environments, mobile robots must contend with vast amounts of continually evolving data. The exploring robot must adapt to its environment and refine its workspace representation with new observations. The key competency we seek is “introspection”: to ability to determine what is perplexing, which can further drive active information acquisition or human disambiguation. The talk explores this in the context of place recognition and semantic mapping.

First, we present “Fabmap 3D”, a new probabilistic framework for appearance-based topological mapping incorporating spatial and visual appearance. Locations are encoded probabilistically as random graphs possessing latent distributions over visual features and pair-wise euclidean distances generating observations modeled as 3D constellations of features observed via noisy range and visual detectors.

Next, we present an approach for continually improving place recognition performance through introspection and then targeted data retrieval. We introduce a dynamic sampling set, the onboard workspace representation that adapts with increasing visual experience. With a topic-based probabilistic model for images, a measure of perplexity evaluates how well a working set of background images explains the online view of the world. Offline, the robot then searches an external resource to seek additional imagery that bolsters its ability to localise next.

Finally, we consider the semantic mapping task, where labels are generated from a variety of classification frameworks, often a precursor for action/decision selection. We introduce and motivate the importance of a classifier’s “introspective capacity”: the ability to mitigate potentially overconfident classifications by a realistic assessment of how qualified the system is to make a judgement on the current test datum. We explore the benefits of an introspective classifier for classification, detection and active learning, in the autonomous driving context.

Bio:
Rohan obtained B. Tech and M. Tech degrees in Computer Science and Engineering from the Indian Institute of Technology (IIT), Delhi, India. He was awarded the Rhodes Scholarship to pursue a DPhil at the Mobile Robotics Group at Oxford University under Prof. Paul Newman (2008-2012). He served as a Research Intern at the Max Planck Institute of Biological Cybernetics (May-June 2007) and as a Visiting Researcher at the Computer Science and Artificial Intelligence Laboratory, MIT (Feb-March 2012).

Rohan’s research interests include appearance-based topological mapping, life-long learning and active information acquisition. He received the Best Conference Paper Award at the International Conference on Intelligent Robots and Systems (IROS) 2013 and was the Best Vision Paper Finalist in the International Conference on Robotics and Automation (ICRA) 2010. He is also the recipient of the National Award by the Indian National Academy of Engineering (INAE) in 2008 and the Best Industry Relevant Thesis by IIT Delhi, 2008.

Apart from Robotics, Rohan works in the area of developing affordable assistive devices for persons with visual impairment, particularly aimed for mobility and education. He has developed devices that are reaching end-users nationally in India and played a lead role in establishing the Assistive Technologies Group at IIT Delhi.

Research Areas:

Impact Areas:

This event is not part of a series.

Created by Nick Roy at Wednesday, April 09, 2014 at 7:37 AM.