Results

Walking Speed A dataset of eleven experiments collected on four different hill sections

tests the walking speed behavior, as summarized in Table 4.1. Overall, the robot climbed around half a kilometer24 _{(461.8 meters, or 810.2 body lengths) of hilly terrain while en-}

countering 111 obstacles and successfully avoiding 107 of them. 49 of the avoided obstacles were detectable by its sensor (trees, tall bushes and walls), and 58 of them were not de- tectable (short bushes, fallen branches and logs). The 4 obstacles the robot failed to avoid were not detectable by its sensor. In other words, the steepest ascent controller, with no obstacle avoidance term introduced, would otherwise have failed to avoid and likely become entrapped by 49 additional obstacles that the robot was able to avoid in these tests. Trials 6 through 10 ended in a summit and Trials 1, 4 and 11 ended when the robot reached an edge of the course (an artificial boundary picked by the operator beyond which it is not safe for the robot). Trials 2, 3 and 5 were stopped when the robot suffered a fault— i.e., it got stuck on a small but rigid branch from the under brush—a non-detectable but insurmountable obstacle. These failures of the world model (specifically, the assumption of Section 3.1.1that any obstacle unseen by the sensor is surmountable), could be addressed by improvements in sensing or locomotion primitives that lie beyond the scope of this paper. Trial 2 contained an intermittent fault where a thick branch trapped the left rear leg. The operator pulled this branch off and the trial continued. Of note, Trial 11 tested the limits of

24_{The total runtime for all walking speed experiments is 37 minutes and 8 seconds. The distance reported} is the direct distance between initial and final locations. The length of the path the robot traversed is not available.

# Location Description Direct Distance Hill Slope Runtime D. O. N. O. Faults Finish

(meters) (degrees) (min:sec)

1 Ridley Creek Medium Forest 69.5 10−15 5:22 8 11 - Edge 2 Ridley Creek Medium Forest 62.3 3−15 6:14 4 12 2x N.O. Fault 3 Ridley Creek Medium Forest 62.9 6−15 4:55 3 13 N.O. Fault

4 Ridley Creek Medium Forest 96.8 3−15 7:31 11 19 - Edge

5 Ridley Creek Steep Forest 18.7 15−18 1:47 2 3 N.O. Fault 6 Penn Park Medium Grassy 27.5 3−12 1:54 1 0 - Summit 7 Penn Park Medium Grassy 20.9 3−20 1:36 2 0 - Summit

8 Penn Park Steep Grassy 36.2 3−33 2:34 7 0 - Summit

9 Penn Park Steep Grassy 22.6 3−33 1:46 4 0 - Summit 10 Penn Park Medium Grassy 28.8 3−20 2:17 7 1 - Summit 11 Penn Park Steep Grassy 15.6 15−36 1:12 0 0 - Edge Table 4.1: Eleven outdoor hill climbing behavior trials including 49 detectable obstacles (D.O.) successfully avoided and 58 non-detectable obstacles (N.O.) successfully mechanically traversed over around half a kilometer of climbing with only 4 faults.

the hill ascent controller, where the hill incline angle reached 36◦_{. Yet, the robot successfully}

traversed this patch of hill and reached the edge.

Running Speed A dataset of nine experiments collected on three of the same hill sec-

tions as the walking speed experiments tests the running speed behavior, as summarized in Table4.2. Overall, the robot climbed 357.8 meters25_{(or 627.7 body lengths) of hilly terrain}

while encountering 89 obstacles, and successfully avoiding 85 of them. 41 of the avoided obstacles were detectable by its sensor (trees, tall bushes and walls), and 44 of them were not detectable (short bushes, fallen branches and logs). The robot got entrapped by 4 ob- stacles. Twice it got caught up on (non-detectable) rigid branches; another two times the navigation failed to clear 2 detectable obstacles, both of which violated the world model in a manner detailed below. In other words, the steepest ascent controller, with no obstacle avoidance term introduced, would otherwise have hit and likely become entrapped by 41 additional obstacles that the robot was able to avoid in these tests.

Trials 15 through 18 ended in a summit and Trials 13 and 20 ended when the robot reached

25

In total, running speed experiments took only 11 minutes and 22 seconds. The distance reported is the direct distance between initial and final locations. The length of the path the robot traversed is not available.

# Location Description Direct Distance Hill Slope Runtime D. O. N. O. Faults Finish (meters) (degrees) (min:sec)

12 Ridley Creek Medium Forest 58.7 10−_{15 1:43 3 11 N.O., W.M.V. Fault}

13 Ridley Creek Medium Forest 87.3 3−_{15 2:55 7 17 - Edge}

14 Ridley Creek Medium Forest 89.5 3−_{15 3:18 12 16 W.M.V, N.O. Fault}

15 Penn Park Medium Grassy 19.3 3−_{12 0:19 1 0 - Summit}

16 Penn Park Medium Grassy 19.7 3−20 0:22 2 0 - Summit

17 Penn Park Steep Grassy 19.3 3−33 0:24 3 0 - Summit

18 Penn Park Steep Grassy 19.1 3−33 0:49 5 0 - Summit

19 Penn Park Medium Grassy 24.2 3−20 0:58 6 0 Hardware Fault

20 Penn Park Steep Grassy 12.5 15−36 0:34 2 0 2x Flip Edge Table 4.2: Nine outdoor hill climbing behavior trials including 41 detectable obstacles (D.O.) successfully avoided and 44 non-detectable obstacles (N.O.) successfully mechanically traversed over around 350 meters of climbing with only 4 obstacle interaction based faults. 2 of these occurred due to robot failure over non-detectable obstacles. The other 2 occurred due to world model violations (W.M.V.) where a complex set of obstacles resulted in the robot control strategy failure.

an edge of the course (see the previous section for details of this termination condition). Trials 12, 14 and 19 were stopped when the robot incurred a fault after the reported distance had been covered. Trial 12 was terminated when the robot reached a fallen trunk and failed to walk around it (which could be considered and edge). This trial also contained an intermittent fault where a thick branch trapped left middle leg. The operator pulled this branch off and the trial continued. Trial 14 ended with the robot climbing over a short bush and losing traction as it can be seen on Figure 4.4. In addition, this trial was interrupted when the robot encountered a concave obstacle region formed by a wide tree and a big fallen branch. The operator moved the robot out of the trap and the trial continued. Trial 19 ended with a hardware failure where the left middle leg cracked. Similar to the previous section, Trial 20 tested the limits of the hill ascent controller where the hill incline angle reached 36◦_{. Unlike the walking speed experiment, the robot would have flipped at}

two different instances without any operator intervention. Each of these interventions are marked as faults.

Figure 4.4: An extreme case: small bush trapping the robot at the end of Trial 14.