Research map / Reinforcement learning / Imitation learning
Trust Region Reward Optimization and Proximal Inverse Reward Optimization Algorithm
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TRRO, a minorization-maximization framework giving monotonic increase of expert likelihood for non-adversarial IRL, and PIRO, a practical SAC-based instantiation with adaptive L2 reward-update regularization.
Problem
Adversarial IRL is unstable, and existing non-adversarial energy-based IRL methods lack formal stability guarantees on reward updates.
Approach
Shows canonical non-adversarial IRL methods maximize expert likelihood (equivalently minimize imitation gap), then derives a trust-region-style MM surrogate for reward updates that guarantees monotonic likelihood improvement under exact policy optimization; PIRO approximates it with an L2 reward-difference penalty with adaptive coefficient and SAC policy steps.
Key results
- PIRO achieves 5967.2 on Ant-v4 vs 5382.5 for the best baseline f-IRL (+584.7) and 5643.7 on Walker2d-v4 (+849.0 gain).
- PIRO reaches 25.7 on AntMaze-UMazeDense-v4 vs 35.6 expert and 11.9 for best baseline HyPE.
- PIRO is slightly below the best baseline on Hopper-v4 (3362.0 vs 3535.7 for GAIL) and on average improves reward recovery across most tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PIRO instead of ML-IRL | IRL on Ant-v4 · Ant-v4 | averaged reward: 5382.5 -> 5967.2 | Helped |
| use PIRO instead of ML-IRL | IRL on Humanoid-v4 · Humanoid-v4 | averaged reward: 5573.4 -> 5954.9 | Helped |
| use PIRO instead of GAIL | IRL on Walker2d-v4 · Walker2d-v4 | averaged reward: 4158.1 -> 5643.7 | Helped |
| use PIRO instead of GAIL | IRL on Hopper-v4 · Hopper-v4 | averaged reward: 3535.7 -> 3362.0 | Hurt |
| use PIRO instead of ML-IRL | IRL on Halfcheetah-v4 · Halfcheetah-v4 | averaged reward: 11873.2 -> 12587.4 | Helped |
| use PIRO instead of HyPE | IRL on AntMaze-UMazeDense-v4 · AntMaze-UMazeDense-v4 | averaged reward: 11.9 -> 25.7 | Helped |
| use PIRO instead of IQ | IRL on AntMaze-MediumDense-v4 · AntMaze-MediumDense-v4 | averaged reward: 3.4 -> 9.4 | Helped |
| use PIRO instead of AIRL | IRL on AntMaze-LargeDense-v4 · AntMaze-LargeDense-v4 | averaged reward: 3.4 -> 8.8 | Helped |
| use PIRO instead of BC | imitation on AdroitHandePen-Human-v1 · AdroitHandePen-Human-v1 | averaged reward: 44.1 -> 254.0 | Helped |
Surprising findings
- Hopper-v4 is the one MuJoCo task where PIRO trails the best baseline (3362.0 vs GAIL's 3535.7 with the table's bolding).
- The theoretical constant C is too large in practice, so the paper replaces it with an adaptive coefficient mu.
Limitations
- Monotonic guarantee assumes exact policy optimization, which is not achieved in practice.
- On-policy sampling may reduce sample efficiency.
- Overall stability depends on the underlying policy optimizer in high-dimensional settings.
Experimental setup
Datasets: MuJoCo expert demonstrations (from f-IRL and ML-IRL) Minari Offline RL datasets (Gym-Robotics) real-world meerkat behavior dataset
Benchmarks: MuJoCo Ant-v4 Humanoid-v4 Walker2d-v4 Hopper-v4 Halfcheetah-v4 AntMaze-UMazeDense-v4 AntMaze-MediumDense-v4 AntMaze-LargeDense-v4 AdroitHandPen-Human-v1 LunarLander 7x7 grid world
Metrics: averaged reward learning curves sample efficiency runtime per iteration
Baselines: BC GAIL MM AIRL FILTER HyPE DAC IQ-Learn ML-IRL f-IRL CSIL P2IL
Builds on: ML-IRL Soft Actor-Critic TRPO SQIL f-IRL
Closest papers by meaning
- Non-Adversarial Inverse Reinforcement Learning via Successor Feature Matching (ICLR 2025 · similarity 0.91)
- Is Inverse Reinforcement Learning Harder than Standard Reinforcement Learning? A Theoretical Perspective (ICML 2024 · similarity 0.90)
- Recent Advancements in Inverse Reinforcement Learning (AAAI 2024 · similarity 0.90)
- Offline Inverse RL: New Solution Concepts and Provably Efficient Algorithms (ICML 2024 · similarity 0.90)
- Rethinking Inverse Reinforcement Learning: from Data Alignment to Task Alignment (NeurIPS 2024 · similarity 0.88)
- A Geometric Perspective on Reward Function Updates in Inverse Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
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.