A Self-Supervised Terrain Roughness Estimator for Off-Road Autonomous Driving

1. A Self-Supervised Terrain Roughness Estimator for Off-Road Autonomous Driving David Stavens and Sebastian Thrun Stanford Artificial Intelligence Lab

2. Self-Supervised Learning “Combines” strengths of multiple sensors. Ultra-Precise, No Range Precise, Long Range

3. Overview • Introduction and Motivation • Classifying Terrain Roughness • Self-Supervised Learning • Experimental Results

4. 2005 DARPA Grand Challenge

5. Velocity Planning for DGC 2005 • Mobile robotics traditionally focuses on steering. • But speed is also important. • Beyond stopping distance and lateral maneuverability. • For Grand Challenge 2005, our vehicle adapted its speed to terrain conditions, minimizing shock: • Increases electrical and mechanical reliability. • Mitigates pose error for laser projection. • Increases traction for improved maneuvers. • Seems to be correlated with slowing on “hard” terrain.

6. Velocity Planning for DGC 2005 • Simple three state algorithm: • Drive at speed limit until shock threshold exceeded. • Slow to bring the vehicle within the shock threshold. • Uses approx. linear relationship between shock and speed. • Which is also important for the new work we present. • Accelerate back to the speed limit. • Discontinuous control problem. • Hard to solve with conventional control approaches. • We used supervised learning.

7. Experiments for DGC 05

8. This Talk: Next Logical Step • We expand our online approach to be proactive. • Our previous approach was entirely reactive. • Difficult to be that precise with laser scanners. • Hence problems of uncertainty and learning. • Accuracy required for roughness detection exceeds that required for obstacle avoidance. • 15cm vs. 2-4cm

9. Other Approaches to Velocity Control • Terramechanics: guidance through rough terrain. • Online assessment only at low speeds. • High speeds require a priori maps. • Our approach is both online and at high speeds. • Speeds up to 35 mph.

10. CMU’s Preplanning Trailer

11. Overview • Introduction and Motivation • Classifying Terrain Roughness • Self-Supervised Learning • Experimental Results

12. Acquiring a 3D Point Cloud

13. Errors in Pose and Projection

14. Z Error vs. Time

15. More than t • “Spread” of plot implies more factors than t. • t is also related to: • Amount/rate of pitching. • Distance between the two scans.

16. Comparing Two Laser Points pair = 1| z |2 – 3| t |4 – 5| xy distance |6 – 7| dpitch1|8 – 7| dpitch2|8 – 9| droll1|10 – 9| droll2|10 • Seven Features: z, t, xy distance, dpitches, drolls • 10 Parameters:1 2 …10 (generated with self-supervised learning)

17. Combining Multiple Comparisons • n pairs in ascending order. • Use weighting because resolution of discontinuities is near resolution of laser. There are not many witness pairs. n R =  pair 11i i = 0 • This generates a score, R, for that patch of terrain. • But how do we assign target values to R?

18. Overview • Introduction and Motivation • Classifying Terrain Roughness • Self-Supervised Learning • Experimental Results

19. Self-Supervised Learning Actual shock when driving over terrain modifies belief about original laser scan. Improves classifier for subsequent scans!

20. Caveat: Must Correct for Speed

21. Mapping from R to Shock Learn a simple suspension model in parallel with the classifier: Rcombined = Rleft 12 + Rright 12 Rleft and Rright is for the terrain under each wheel.

22. Overview • Introduction and Motivation • Classifying Terrain Roughness • Self-Supervised Learning • Experimental Results

25. Summary • Road shock provides ground truth for previously perceived patches of road. • Perception model improves in real-time. • Future terrain assessment is more precise. • A faster route completion time is possible. • For the same amount of shock. • Works either “offline” or “as you drive.” • Offline results presented.

26. Questions?