Loading presentation...

Present Remotely

Send the link below via email or IM


Present to your audience

Start remote presentation

  • Invited audience members will follow you as you navigate and present
  • People invited to a presentation do not need a Prezi account
  • This link expires 10 minutes after you close the presentation
  • A maximum of 30 users can follow your presentation
  • Learn more about this feature in our knowledge base article

Do you really want to delete this prezi?

Neither you, nor the coeditors you shared it with will be able to recover it again.


ICRA2013 DDF-SAM 2.0 Interactive Presentation

This presentation was created as an interactive presentation for the 2013 IEEE International Conference on Robotics and Automation (ICRA 2013), and shown on May 9th, 2013 in Karlsruhe, Germany. See alexcunningham.net/publications for original paper.

Alex Cunningham

on 10 May 2013

Comments (0)

Please log in to add your comment.

Report abuse

Transcript of ICRA2013 DDF-SAM 2.0 Interactive Presentation

Approximates joint over shared variables as product of marginals
Provably conservative
Computes marginals directly from Bayes tree using efficient caching technique Alex Cunningham, Vadim Indelman and Frank Dellaert
Center for Robotics and Intelligent Machines @ Georgia Tech
Atlanta, GA, USA DDF-SAM 2.0: Consistent Distributed Smoothing and Mapping Summarized Map Eliminate
all poses Eliminate one pose Goal: Joint marginal density on shared variables
Start with
Eliminate non-shared variables
Remaining factor on shared variables
Complication: must linearize system before elimination Basic Summarization Disaster City Search and Rescue Training Center Hazards Small and light enough for collapsed spaces
Send robots instead of people Why Robots? Goals
Teams of robots to explore disaster sites and battlefields
Provide situational awareness to users
Use autonomous exploration to make system more robust to failure Measurements Poses Landmarks Estimate trajectory and map
Landmarks extracted from sensor measurements Simultaneous Localization and Mapping (SLAM) E D C B Neighborhood of A A Pure local mapping Goal: extend sensor horizon of each robot using neighborhood information Each robot performs local SLAM
Choose a subset of variables to share
Usually landmarks
Summarize local maps for distribution
Robots cache and fuse information from the neighborhood DDF-SAM Approach Motivation: Multi-robot Exploration
in Dangerous Environments Key Challenge: Robust Perception Multi-robot SLAM Example: 3 Robot Network Robot communication
neighborhood Decentralized Data Fusion (DDF) Requirements Robust to robot platform failures
Robust to communication failures
Minimize computational load on individual robots
Minimize communication bandwidth Factor Graph Representation Graphs have variables and factors
Variables are landmarks and poses
Factors represent measurements
Casts inference as optimization DDF-SAM 1.0: Separate Local and Neighborhood Graphs Evaluation Results: Timing vs. Summarization Method Local Update Summarization Neighborhood Update Simulated Aerial Visual SLAM
3 Robots
Randomly generated landmarks
Known data associations
Known reference frame (e.g., GPS)
Shared linearization points Results: Consistency Compare information content of shared summarized maps
Uses batch summarization as standard of comparison
Approximate technique remains conservative Results: Augmented Local Map Estimates Landmark labeling
Red: pure local
Green: pure neighborhood
Blue: overlapped landmarks Performance Conclusions
Batch summarization on local system remains fastest
Summarization has a significant effect on update timing
Naive Bayes approximate summarization significantly improves sparsity Implementation
GTSAM for graph operations
MATLAB wrapper for visualization DDF-SAM 2.0 Augmented Local Map - combines local and neighborhood information in incremental solver
Antifactor Downdating - exactly subtracts double-counted information
Exact and Approximate Summarization Techniques Challenge: Information Double-counting Full Naive System Single-robot Data Structure Conservative approach to double-counting avoidance
Information moves from local to neighborhood graph
Rebuilds neighborhood graph from cached summarized maps
Employs constrained optimization to merge maps Shameless Plug:
GTSAM 2.3 Release! Download: tinyurl.com/gtsam Antifactor Downdating Augmented Local Map And Antifactors Neighborhood Update Incorporate received summarized maps
Summarized maps are just fragments of factor graphs
Can add directly to iSAM solver Sharing Summarized Map Summarize directly from Bayes tree
Doing this naively double-counts information
We add antifactors to cancel contributions of neighboring robots Goal: Subtract Information Adding new measurements to an information-form system is additive
Subtracting information matrix corresponding to double-counted information is possible The Antifactor Is the negation of a factor, which can be added to an iSAM solver to subtract specific contributions Exact and Approximate Summarization from the Bayes Tree Schur Complement Reordering Naive Bayes Summarization Remove contribution: Choose elimination ordering to make summarization trivial
Place shared variables at end of ordering
Equivalent to placing variables at root of Bayes tree Contributions Measurement Prediction: Single factor: Factor graph: Cyclic Network Graphs Unless information sources are isolated, contributions from different sources become difficult to distinguish
Double-counting measurements results in overconfident estimates
The full solution becomes inconsistent Simple Solution: Change the Network Graph Double-counting will not occur in a DAG structure
However, reduces resilience to changes in network graph Goal: Opportunistic message passing Conclusions and Future Work It is possible to combine local and neighborhood information in a consistent manner
Antifactors provide a means to exactly subtract information Future Work Relax assumptions on GPS and shared linearization points
Evaluate tree-subgraph approximations for summarization
Send smaller partitions of summarized maps
Adaptively choose neighbors change
Full transcript