2024
Manipulating Neural Path Planners via Slight Perturbations got accepted by RAL
Alert! A large set of malicious behaviors can be injected and triggered by slightly perturbing the neural path planner’s input. Check out our latest work on backdoor attacks on neural path planners.
DSCRL got accepted by ICRA 24’
Want to effectively harness learning-based navigation and control tasks with programmable logic specifications? Check out Differentiable Specification Constrained Reinforcement Learning (DSCRL), a logic-constrained, map-adaptive co-learning framework for navigation/planning and control.
2023
Two autonomous driving planning patents filed with Baidu are published!
Two patents filed with Baidu are published! Congrats to the colleagues at Baidu!
Submitted a Paper about Programmable Backdoor Attack
Our formalization allows an attacker to precisely characterize the behavior they intend the planner to execute.
Published a new pre-print on programmable deep reinforcement learning
Our framework uses formal signal temporal logic’s differentiability for sample-efficient constrained reinforcement learning. See all trained agents here.
2022
Presented at IROS 22’
Presentation video of Model-free Neural Lyapunov Control for Safe Robot Navigation.
Presented at ECML-PKDD 22’
Pitch video
IROS Late Breaking Results Poster
A follow-up work for my internship at Baidu A Learning-Enhanced Parameter Tuner Improves Motion Planner Deployment Scalability in Autonomous Driving got accepted as IROS Late Breaking Results Poster. We enhanced our previous critic model with VectNet.
IROS 22’ Accepted Our Work
Our work Model-free Neural Lyapunov Control for Safe Robot Navigation got accepted at IROS 22’!
ECML-PKDD 22’ Accepted Our Work
Our work Defending Observation Attacks in Deep Reinforcement Learning via Detection and Denoising got accepted at ECML-PKDD 22’!
Published a new pre-print on safe reinforcement learning
Our approach involves learning a neural Lyapunov function within the reinforcement learning loop and using it to guarantee safety after deployment.