Research map / Reinforcement learning / Offline RL
RAMBO-RL: Robust Adversarial Model-Based Offline Reinforcement Learning
Marc Rigter, Bruno Lacerda, Nick Hawes
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Add adversarial model updates to RAMBO (ablation) | RAMBO · MuJoCo-v2 | normalized return: 694.4 ± 56.5 -> 826.2 ± 33.8 | Improved |
| Add adversarial model updates to RAMBO (ablation) | RAMBO · AntMaze-v0 | normalized return: 45.8 ± 21.8 -> 67.0 ± 14.9 | Improved |
| Choose hyperparameters offline using the Q-value magnitude and stability heuristic instead of online tuning | RAMBO · MuJoCo-v2 | normalized return: 826.2 ± 33.8 -> 785.5 ± 40.4 | Hurt |
| Choose hyperparameters offline using the Q-value magnitude and stability heuristic instead of online tuning | RAMBO · AntMaze-v0 | normalized return: 67.0 ± 14.9 -> 37.8 ± 12.4 | Hurt |
| Use adversarial model training instead of COMBO's value-function pessimism | RAMBO vs COMBO · Single Transition Example | performance: 1.41 ± 0.12 -> 1.49 ± 0.05 | Improved |
| Use adversarial weighting λ = 0 instead of λ = 3e-4 (ablation) | RAMBO · MuJoCo-v2 Random HalfCheetah | normalized return: 34.5 ± 3.7 -> 40.0 ± 2.3 | Mixed |
| Use adversarial weighting λ = 0 instead of λ = 3e-4 (ablation) | RAMBO · MuJoCo-v2 Medium HalfCheetah | normalized return: 77.6 ± 1.5 -> 75.5 ± 6.0 | Hurt |
| Use rollout length k = 5 instead of k = 2, with λ = 3e-4 (ablation) | RAMBO · MuJoCo-v2 Medium HalfCheetah | normalized return: 72.5 ± 4.4 -> 77.6 ± 1.5 | Improved |
| Use rollout length k = 5 instead of k = 2, with λ = 3e-4 (ablation) | RAMBO · MuJoCo-v2 Random Walker2D | normalized return: − -> 0.2 ± 0.5 | Mixed |
| Use behavior-cloning policy initialization (ablation) | RAMBO · MuJoCo-v2 | normalized return | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- COMBO — COMBO: Conservative Offline Model-Based Policy Optimization (NeurIPS 2021) · better on 2 of 2 shared results
- MOReL — MOReL: Model-Based Offline Reinforcement Learning (NeurIPS 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Off-Policy Actor-Critic for Adversarial Observation Robustness: Virtual Alternative Training via Symmetric Policy Evaluation (ICML 2025 · similarity 0.90)
- Adversarially Trained Weighted Actor-Critic for Safe Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- Adversarial Diffusion for Robust Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Making Offline RL Online: Collaborative World Models for Offline Visual Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- Adversarial Policy Optimization for Offline Preference-based Reinforcement Learning (ICLR 2025 · similarity 0.89)
- Scaling Offline Model-Based RL via Jointly-Optimized World-Action Model Pretraining (ICLR 2025 · 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.
