IIIT Hyderabad Publications |
|||||||||
|
Talk to the Vehicle: Language Conditioned Autonomous Navigation of Self Driving CarsAuthors: Sriram N N,Tirth Maniar,Jayaganesh Kalyanasundaram,Vineet Gandhi,Madhava Krishna Conference: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2019 2019) Location The Venetian Macao, Macau, China Date: 2019-11-04 Report no: IIIT/TR/2019/75 AbstractWe propose a novel pipeline that blends encodingsfrom natural language and 3D semantic maps obtained fromvisual imagery to generate local trajectories that are executedby a low-level controller. The pipeline precludes the need fora prior registered map through a local waypoint generatorneural network. The waypoint generator network (WGN) mapssemantics and natural language encodings (NLE) to localwaypoints. A local planner then generates a trajectory fromthe ego location of the vehicle (an outdoor car in this case) tothese locally generated waypoints while a low-level controllerexecutes these plans faithfully. The efficacy of the pipelineis verified in the CARLA simulator environment as well ason local semantic maps built from real-world KITTI dataset.In both these environments (simulated and real-world) weshow the ability of the WGN to generate waypoints accuratelyby mapping NLE of varying sequence lengths and levels ofcomplexity. We compare with baseline approaches and showsignificant performance gain over them. And finally, we showreal implementations on our electric car verifying that thepipeline lends itself to practical and tangible realizations inuncontrolled outdoor settings. In loop execution of the proposedpipeline that involves repetitive invocations of the network iscritical for any such language-based navigation framework.This effort successfully accomplishes this thereby bypassingthe need for prior metric maps or strategies for metric levellocalization during traversal. Full paper: pdf Centre for Robotics |
||||||||
Copyright © 2009 - IIIT Hyderabad. All Rights Reserved. |