Friday, April 15, 2011

Large-scale Multi-robot Mapping in MAGIC 2010

5th IEEE Conference on Robotics, Automation and Mechatronics (RAM 2011)

Authors: Robert Reid, Thomas Bräunl

Abstract: We describe a large-scale decentralised multi-robot mapping system that outputs globally optimised metric maps in real-time. The mapping system was used by team WAMbot in the finals of the Multi-Autonomous Ground-robotics International Challenge (MAGIC 2010). Research contributions include a novel large-scale multi-robot graph-based non-linear map optimisation approach, a hybrid decentralised and distributed mapping system and novel graphics processing unit (GPU) based approaches for accelerating intensive map matching and fusion operations. Our mapping system scales linearly with map size and on commodity hardware can easily map a 500m×500m urban area. We demonstrate robust, highly efficient and accurate mapping results from two different fleets of mobile robots. Videos, maps and timing results from the MAGIC 2010 challenge are presented.

Index Terms: Large-scale multi-robot mapping, simultaneous localisation & mapping (SLAM), graph-based SLAM, distributed, decentralised, GPU map fusion, GPU map correlation.

Videos: These videos present the output of the Mapbuilder system and play at 15 × real-time speed. They are best viewed full-screen and at 1920×1080 (click the full-screen icon in the lower right of the video and also select 1080p). The global occupancy maps are a top-down view showing obstacles, free-space and unknown areas in the environment. The submap spatial constraints are shown as blue lines. We note that tunable parameters in the mapping system were not changed when moving between environments.

MAGIC 2010 Challenge Phase 2: This video shows 7 UGVs exploring an agricultural show-ground, passing in and out of horse stables. Figure 5 in the conference paper shows the final map overlaid on aerial imagery. The 200m× 160m map is metrically accurate and correctly georeferenced by the intermittent and noisy GPS data. To quantify the RMS error in the final map we selected a set of evenly distributed features in the map and measured their linear error to the aerial image. The estimated mean linear error was 0.57m. The final map has 446 submaps, 419 ground-truth constraints and 1284 submap constraints. For the final map each execution cycle it took 21ms to incrementally optimise the entire pose-graph, and 87ms to build and output the global occupancy map. The initial submaps take some time to become connected due to occlusions and ambiguous geometry in the matching process, however after driving several meters each UGV correctly joins its current submap to the global graph. Slow drifts in submaps are observed as transient GPS noise is filtered into ground-truth data. If additional ground truth constraints had been supplied by the operator, spatial errors could have been reduced by an order of magnitude.

MAGIC 2010 Old Ram Shed Challenge: This video shows 5 UGVs exploring a 70m ×40m agricultural shed. Again the submaps take some time to become connected however once the UGVs begin moving the map is rapidly constrained and is metrically accurate. Note that barriers inside the shed were temporary and very few walls and angles were regular. No ground-truth was made available to evaluate the accuracy of this map. The final map has 260 submaps and 2170 submap constraints. For the complete map each execution cycle took 23ms to incrementally optimise the pose graph and 85ms to build and output the global occupancy map.

No comments:

Post a Comment