Rethinking Robustness Assessment: Adversarial Attacks on Learning-based Quadrupedal Locomotion Controllers

Fan Shi *, Chong Zhang *, Takahiro Miki, Joonho Lee, Marco Hutter, Stelian Coros
Robotic Systems Lab, ETH Zurich; Computational Robotics Lab, ETH Zurich; ETH AI Center; National University of Singapore

Accepted by RSS 2024 . pdf

Key takeaways:
1. Domain randomization is NOT sufficient for robustness
2. Multi-modality attacks are much stronger
3. Adversaries with exteroception can smartly leverage terrains
4. AI beats human robot experts on attack strategies
5. SOTA robust control policy, which won DARPA SubT Challenge, is still fragile against even mild attack sequences
6. Control weakness can be covered by training together with adversarial attacks





Supplement videos:

   

- - -