Off-Road Navigation With End-to-end Imitation Learning for Continuously Parameterized Control

End-to-End Learning for Visual Navigation of Forest Environments

“…Weight of Rovers (kg) 5: Highways and traffic road [44] Racing track on traffic road 210 1231 [40] Highways (sunlight facing the camera, high contrast sunlight, shadows, covered in snow) 14,000 1579 [37] Traffic road 380 - [38] Traffic road and walkways in parks 8400 -4: Off-road on cemented paths, short grass, pebbles, dirt and dry leaves [41] Off-road racing track 300 22 [42] Mowed and short grass off-trail 380 35 [43] Cemented and off-road trails with pebbles, dirt, sand, grass and fallen leaves, with few obstacles 5500 35 3: Sidewalks and walkways in urban environment [47] Static environments: walkways in office areas, laboratory space and corridors; dynamic environments: sidewalks among crowds. 355 12 [45] Paved road cemented on grass 800 50 [46] Mowed lawn, short grass, and trees in urban environment 1000 17 [53] Sidewalks outside malls and office buildings 5500 - [48] Walkways in neighborhoods and parks 60 5 [39] Parking lots, city roads and sidewalks 60 62 [25] Corridor indoors and stone trail outdoors 400 30…”

“…In outdoor environments, algorithms using end-to-end learning have demonstrated some promising results in autonomous driving on well-paved roads in structured urban environments under varying lighting and weather conditions [37][38][39][40]. Moreover, recently, a few studies have investigated endto-end learning algorithms for off-road navigation [41][42][43]. For example, using control policy predictions of steering and throttle commands trained using an end-to-end imitation learning approach, a 1/5-scaled RC vehicle was able to successfully navigate a dirt trackwithout obstacles-at high speed [41].…”