Research map / Reinforcement learning / Imitation learning
BC-IRL: Learning Generalizable Reward Functions from Demonstrations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace the MaxEnt reward-separation objective with an outer behavioral-cloning loss differentiated through policy optimization | BC-IRL-PPO versus MaxEnt IRL baselines · Point mass navigation, unseen start distribution | distance to goal: 0.04 { $\pm$ 0.01 } versus 0.53 { $\pm$ 0.78 } for AIRL | Helped |
| Learn a reward with the BC-IRL-PPO procedure rather than AIRL | Policy trained from the learned reward · Fetch Reach, hard goal/start distribution | success rate: 0.34 -> 0.76 | Helped |
| Transfer the learned reward and train a new policy, rather than transfer only the learned policy | BC-IRL-PPO · Fetch Reach, hard generalization | success rate: 0.35 -> 0.76 | Helped |
| Transfer and fine-tune both policy and reward instead of transferring only the reward | BC-IRL-PPO · Fetch Reach, hard start distribution | success rate: 0.76 -> 0.80 | Helped |
| Use gradient-based MPC rather than PPO in the inner loop (ablation) | BC-IRL · Obstacle point mass navigation | outer-loop objective optimization | Hurt |
| Vary the number of inner-loop policy optimization steps K (ablation) | BC-IRL-PPO · Point mass navigation without obstacle | training performance | No clear effect |
| Reduce demonstrations from 100 to 5 (ablation) | BC-IRL-PPO · Fetch Reach, +75% noise setting | success rate: 1.00 -> 1.00 | No clear effect |
| Reduce demonstrations from 100 to 5 (ablation) | BC-IRL-PPO · Fetch Reach, +100% noise setting | success rate: 0.76 -> 0.69 | Hurt |
| Reduce demonstrations from 100 to 5 (ablation) | AIRL · Fetch Reach, +100% noise setting | success rate: 0.38 -> 0.42 | Helped |
| Increase the policy learning rate to 1e-2 or 1e-1 (ablation) | BC-IRL-PPO · Fetch Reach training setting | success rate: 0.97 { $\pm$ 0.03 } -> 0.14 { $\pm$ 0.23 } | Hurt |
| Increase the reward learning rate (ablation) | BC-IRL-PPO · Fetch Reach training setting | success rate: 0.96 { $\pm$ 0.01 } -> 0.71 { $\pm$ 0.45 } | Hurt |
| Use the BC-IRL learned reward instead of a sparse hand-coded reward (ablation) | Policy trained on obstacle point mass task · Obstacle point mass navigation | distance to goal | Helped |
Closest papers by meaning
- Compatible Reward Inverse Reinforcement Learning (NeurIPS 2017 · similarity 0.93)
- Is Inverse Reinforcement Learning Harder than Standard Reinforcement Learning? A Theoretical Perspective (ICML 2024 · similarity 0.92)
- Recent Advancements in Inverse Reinforcement Learning (AAAI 2024 · similarity 0.92)
- Offline Inverse RL: New Solution Concepts and Provably Efficient Algorithms (ICML 2024 · similarity 0.91)
- Towards the Transferability of Rewards Recovered via Regularized Inverse Reinforcement Learning (NeurIPS 2024 · similarity 0.90)
- A Geometric Perspective on Reward Function Updates in Inverse Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
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.