Rethinking Robustness Assessment: Adversarial Attack on Learning-based Quadruped Locomotion Control

In our RSS 2024 paper, we present a novel adversarial attack method designed to identify failuer cases in any type of locomotion controller, including state-of-the-art reinforcement learning (RL)-based controllers. Traditional heuristic tests, such as standard benchmarks or human experience, often fall short in uncovering these vulnerabilities. Our approach reveals the vulnerabilities of black-box neural network controllers, providing valuable insights that can be leveraged to enhance robustness through retraining.
Project website: fanshi14.github.io/me/rss24.html
Paper link: arxiv.org/abs/2405.12424

[@AnnanTang]

Super nice work!🎉

[@mccanlessdesign]

Super, super work!

[@tommyliao5508]

Amazing work, ppl!

[@liangcherry]

Cool Work!

[@monkeysfromvenus]

this is so cool

[@dter706]

What are the benefits of the RL-based controller over say a robust MPC?

[@user-es2zs4br2k]

Handling model mismatch (e.g. unknown payload), nonlinearities, assumption violations (e.g. slippery),
uncertainties (perception noises)

[@broken_symmetry]

@@user-es2zs4br2k the robustness is NOT due to some magic in RL but due to the fact that via offline training the solution map of the robust MPC problem gets stored in the neural network, while traditional MPC solvers can not solve such complex problems online. To me, they’re just different methods to solve the same optimization problem.

[@Lightning---K.]

Fantastic, thanks to you guys it won´t be possible to kick robots so that they fall and smash on the ground ... wait let me rethink ... we will have less chances to destroy a robot with a malfunction on a killing mission ? Future A.I. is going to wipe us out ..one after another ... didn´t you guys learn from the Sci-Fi movies ? Mashines against humans ? Terminator , Matrix, I, Robot ?
At least leave a vulnerable area on the robot ... We live in really interesting times !

[@Atmatan_Kabbaher]

You obviously didn't learn from them: the lesson wasn't not to make AI. It was to treat them like living things and give them place in society.

[@minsikbuff]