Research map / Reinforcement learning / Policy gradient and actor-critic
Langevin Rollout Optimization for Modelic Reinforcement Learning
Tianyi Zhang, Likun Wang, Guojian Zhan, Feihong Zhang, Yang Guan, Yao Lyu, Shengbo Eben Li
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 LaRO (Langevin gradient refinement) in parallel with MPPI via MLAP | BOOM · H-Bench Hand (7 tasks avg) | Total Average Return: 634.2 -> 773.3 | Improved |
| add LaRO via MLAP | BOOM · DMC Suite Hard (7 tasks avg) | Total Average Return: 802.7 -> 847.2 | Improved |
| add LaRO via MLAP | BOOM · H1hand-run | Total Average Return: 329.3 -> 823.0 | Improved |
| use learned score network instead of exact BPTT gradients (ablation) | LaRO · H1hand-run | Total Average Return | No clear effect |
| use learned score network instead of BPTT gradients (ablation) | LaRO · H1hand-run | latency (ms): 12.72 -> 1.79 | Improved |
| lower Langevin temperature alpha (towards deterministic gradient ascent) (ablation) | LaRO · H1hand-run | Total Average Return | Hurt |
| use 3 Langevin steps instead of 5 (ablation) | LaRO · H1hand-run | Total Average Return | Hurt |
| use 8 Langevin steps instead of 5 (ablation) | LaRO · H1hand-run | Total Average Return | Hurt |
| LaRO-only planner versus MPPI-only planner (ablation) | BOOM-L · H1hand-run | Total Average Return | No clear effect |
| combine MPPI and LaRO with max look-ahead selection (MLAP) vs either alone (ablation) | BOOM-L · H1hand-run | Total Average Return | Improved |
| add LaRO and score network on top of BOOM | BOOM-L · Average of 4 tasks | inference latency (ms): 28.95 -> 34.76 | Hurt |
| add LaRO and score network on top of BOOM | BOOM-L · Average of 4 tasks | training time (h): 20.08 -> 25.51 | Hurt |
Closest papers by meaning
- Provable and Practical: Efficient Exploration in Reinforcement Learning via Langevin Monte Carlo (ICLR 2024 · similarity 0.89)
- Langevin Soft Actor-Critic: Efficient Exploration through Uncertainty-Driven Critic Learning (ICLR 2025 · similarity 0.89)
- Mean Field Langevin Actor-Critic: Faster Convergence and Global Optimality beyond Lazy Learning (ICML 2024 · similarity 0.88)
- COPlanner: Plan to Roll Out Conservatively but to Explore Optimistically for Model-Based RL (ICLR 2024 · similarity 0.88)
- TD-MPC2: Scalable, Robust World Models for Continuous Control (ICLR 2024 · similarity 0.88)
- SLOPE: Optimistic Potential Landscape Shaping for Model-based Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
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.
