Research map / Reinforcement learning / Policy gradient and actor-critic
Robust Reinforcement Learning on State Observations with Learned Optimal Adversary
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train an online learned adversary using PPO | attack against vanilla PPO · Hopper, Walker2d, Ant, HalfCheetah | attack reward: 794$\pm$238 -> \bf636$\pm$ 9; 1336 $\pm$ 654 -> \bf1086$\pm$516; 268 $\pm$227 -> \bf-872 $\pm$ 436; 489$\pm$ 758 -> \bf | Hurt |
| add ATLA training with an LSTM policy and state-adversarial regularization | PPO · Hopper | best attack reward: 636$\pm$ 9 -> 1772$\pm$ 802 | Helped |
| add ATLA training with an LSTM policy and state-adversarial regularization | PPO · Walker2d | best attack reward: 1086$\pm$516 -> 3239$\pm$ 894 | Helped |
| add ATLA training with an LSTM policy and state-adversarial regularization | PPO · Ant | best attack reward: -872 $\pm$ 436 -> 3765$\pm$ 101 | Helped |
| add ATLA training with an LSTM policy and state-adversarial regularization | PPO · HalfCheetah | best attack reward: -660$\pm$ 218 -> 4806$\pm$ 603 | Helped |
| add state-adversarial regularization to ATLA-PPO (LSTM) (ablation) | ATLA-PPO (LSTM) · Hopper | best attack reward: 1224$\pm$ 191 -> 1772$\pm$ 802 | Helped |
| add state-adversarial regularization to ATLA-PPO (LSTM) (ablation) | ATLA-PPO (LSTM) · Walker2d | best attack reward: 3219 $\pm$ 1132 -> 3239$\pm$ 894 | No clear effect |
| add state-adversarial regularization to ATLA-PPO (LSTM) (ablation) | ATLA-PPO (LSTM) · Ant | best attack reward: 716$\pm$ 256 -> 3765$\pm$ 101 | Helped |
| add state-adversarial regularization to ATLA-PPO (LSTM) (ablation) | ATLA-PPO (LSTM) · HalfCheetah | best attack reward: 2485$\pm$ 1488 -> 4806$\pm$ 603 | Helped |
| use an LSTM policy instead of an MLP policy (ablation) | ATLA-PPO · Hopper, Walker2d, Ant, HalfCheetah | best attack reward: 976$\pm$ 40 -> 1224$\pm$ 191; 2213$\pm$ 915 -> 3219 $\pm$ 1132; 33$\pm$327 -> 716$\pm$ 256; 2170$\pm$ 2097 -> 2485$\pm$ | Mixed |
| use LSTM policy in vanilla PPO without ATLA (ablation) | PPO · HalfCheetah | best attack reward: -660$\pm$ 218 -> -886$\pm$ 30 | Hurt |
| increase SA-PPO regularization to $\kappa=1.0$ (ablation) | SA-PPO · Figure 3 environment | natural reward: 1436 $\pm$ 96 | Hurt |
Closest papers by meaning
- Beyond Worst-case Attacks: Robust RL with Adaptive Defense via Non-dominated Policies (ICLR 2024 · similarity 0.92)
- Robust Multi-Agent Reinforcement Learning with Stochastic Adversary (ICML 2025 · similarity 0.91)
- Rethinking Adversarial Policies: A Generalized Attack Formulation and Provable Defense in RL (ICLR 2024 · similarity 0.90)
- Off-Policy Actor-Critic for Adversarial Observation Robustness: Virtual Alternative Training via Symmetric Policy Evaluation (ICML 2025 · similarity 0.90)
- Adversarial Diffusion for Robust Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Enhancing Robustness in Deep Reinforcement Learning: A Lyapunov Exponent Approach (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.