Research map / Reinforcement learning / Offline RL
LAPO: Latent-Variable Advantage-Weighted Policy Optimization for Offline Reinforcement Learning
Xi Chen, Ali Ghadirzadeh, Tianhe Yu, Jianhao Wang, Yuan Gao, Wenzhe Li, Liang Bin, Chelsea Finn, Chongjie Zhang
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 |
|---|---|---|---|
| LAPO latent generative model vs PLAS VAE | offline RL policy · Toy navigation task | average return: 1.03 -> 4.24 | Improved |
| replace Gaussian with GMM policy in AWAC | AWAC · Toy navigation task | average return: 2.43 -> 4.39 | Improved |
| replace Gaussian with GMM policy in AWAC | AWAC · Walker2d-mix-forward-v1 | normalized score: 71.89 -> 78.09 | Improved |
| replace AWAC Gaussian with GMM policy | AWAC · Maze2d-umaze-v1 | normalized score: 94.53 -> 19.45 | Hurt |
| use LAPO instead of next best baseline | LAPO · heterogeneous tasks (12) | normalized score: 49% average improvement | Improved |
| use LAPO on narrow/biased locomotion datasets | LAPO · D4RL locomotion random/medium/expert | normalized score | Mixed |
| LAPO vs CQL on Hopper-medium | LAPO · Hopper-medium-v2 | normalized score: 71.59 -> 51.63 | Hurt |
| LAPO vs best baseline on Hopper-random | LAPO · Hopper-random-v2 | normalized score: 8.33 -> 23.46 | Improved |
| remove latent policy training (sample z from prior) (ablation) | LAPO · Walker2D-medium and Walker2d-mix-backward | normalized score | No clear effect |
| remove latent policy training (sample z from prior) (ablation) | LAPO · Maze2D-medium and Antmaze-medium-diverse | normalized score | Hurt |
| remove limit on latent z range [-zmax, zmax] (ablation) | LAPO · four tasks | normalized score | Hurt |
| LAPO vs best baseline on Maze2d-large | LAPO · Maze2d-large-v1 | normalized score: 64.30 -> 200.56 | Improved |
Later papers that compared with this method
2 later accepted papers report LAPO as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Learning on One Mode: Addressing Multi-modality in Offline Reinforcement Learning (ICLR 2025 · better on 1 of 3 shared results)
- Design from Policies: Conservative Test-Time Adaptation for Offline Policy Optimization (NeurIPS 2023 · better on 4 of 6 shared results)
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
- Constrained Latent Action Policies for Model-Based Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.92)
- Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models (NeurIPS 2025 · similarity 0.90)
- Reasoning with Latent Diffusion in Offline Reinforcement Learning (ICLR 2024 · similarity 0.90)
- Adaptive Advantage-Guided Policy Regularization for Offline Reinforcement Learning (ICML 2024 · similarity 0.89)
- GTA: Generative Trajectory Augmentation with Guidance for Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- A2PO: Towards Effective Offline Reinforcement Learning from an Advantage-aware Perspective (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.
