Research map / Reinforcement learning / Policy gradient and actor-critic
Learning Anisotropic Value Geometry with Finsler Reinforcement Learning
Jumman Hossain, Nirmalya Roy
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 CVaR objective to PPO (ablation) | PPO · MuJoCo aggregate (SlopedHopper-12deg, Walker2d-5deg, HalfCheetah-5deg) | CVaR0.1 cost: 1.50 -> 1.20 | Improved |
| add CVaR objective to PPO (ablation) | PPO · MuJoCo aggregate (SlopedHopper-12deg, Walker2d-5deg, HalfCheetah-5deg) | energy: 1.00 -> 1.15 | Hurt |
| add Finsler cost shaping to PPO (ablation) | PPO · MuJoCo aggregate (SlopedHopper-12deg, Walker2d-5deg, HalfCheetah-5deg) | energy: 1.00 -> 0.93 | Improved |
| add Finsler cost shaping to PPO (ablation) | PPO · MuJoCo aggregate (SlopedHopper-12deg, Walker2d-5deg, HalfCheetah-5deg) | CVaR0.1 cost: 1.50 -> 1.25 | Improved |
| combine Finsler cost with CVaR (FiRL) versus Finsler alone (ablation) | PPO + Finsler · MuJoCo aggregate (SlopedHopper-12deg, Walker2d-5deg, HalfCheetah-5deg) | CVaR0.1 cost: 1.25 -> 0.80 | Improved |
| combine Finsler cost with CVaR (FiRL) versus Finsler alone (ablation) | PPO + Finsler · MuJoCo aggregate (SlopedHopper-12deg, Walker2d-5deg, HalfCheetah-5deg) | success rate: 94.6 -> 97.4 | Improved |
| replace Finsler cost with non-convex ANF cost (ablation) | FiRL · RoughTerrain | success rate: 95.2% -> 80.4% | Hurt |
| replace Finsler cost with non-convex ANF cost (ablation) | FiRL · RoughTerrain | energy (CoT): 0.78 -> 0.95 | Hurt |
| use FiRL instead of PPO on 12 degree slope | PPO · Hopper, slope 12 deg | success rate: 82% -> about 98% | Improved |
| lower risk level alpha below 0.05 (ablation) | FiRL · SlopedHopper-12deg | energy cost | Hurt |
| add Gaussian angular noise (sigma=5 deg) to terrain-normal estimate (ablation) | FiRL · MuJoCo noisy terrain | success rate: remains above 94% | No clear effect |
| use FiRL instead of CVaR-PPO | CVaR-PPO · Isaac Sim Staircase | energy: 1.12 -> 0.94 | Improved |
Closest papers by meaning
- QPRL : Learning Optimal Policies with Quasi-Potential Functions for Asymmetric Traversal (ICML 2025 · similarity 0.88)
- Physics-informed Value Learner for Offline Goal-Conditioned Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Reward Penalties on Augmented States for Solving Richly Constrained RL Effectively (AAAI 2024 · similarity 0.84)
- Risk-Conditioned Reinforcement Learning: A Generalized Approach for Adapting to Varying Risk Measures (AAAI 2024 · similarity 0.84)
- Flattening Hierarchies with Policy Bootstrapping (NeurIPS 2025 · similarity 0.84)
- Mirror Descent Beyond Euclidean Stability: An Exponential Separation in Initialization Sensitivity (NeurIPS 2026 · similarity 0.84)
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.
