Research map / Reinforcement learning / Visual and representation RL
Inverse Reinforcement Learning with Natural Language Goals
Li Zhou, Kevin Small
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 |
|---|---|---|---|
| Use variational goal generation with EGR, HGR, and HGS | LangGoalIRL · R2R seen validation | success rate: 0.449 ± 0.0130 -> 0.530 ± 0.0138 | Improved |
| Add self-supervised fine-tuning in the unseen environment | LangGoalIRL · R2R unseen validation | success rate: 0.357 ± 0.0089 -> 0.491 ± 0.0065 | Improved |
| Add attention over natural-language goals | LangGoalIRL_BASE · R2R seen validation | success rate: 0.359 ± 0.0117 -> 0.449 ± 0.0130 | Improved |
| Add EGR alone (ablation) | LangGoalIRL · R2R seen validation | success rate: 0.449 ± 0.0130 -> 0.450 ± 0.0122 | No clear effect |
| Add HGR alone (ablation) | LangGoalIRL · R2R seen validation | success rate: 0.449 ± 0.0130 -> 0.427 ± 0.0048 | Hurt |
| Add HGS alone (ablation) | LangGoalIRL · R2R seen validation | success rate: 0.449 ± 0.0130 -> 0.399 ± 0.0124 | Hurt |
| Add HGR to EGR (ablation) | LangGoalIRL · R2R seen validation | success rate: 0.450 ± 0.0122 -> 0.504 ± 0.0185 | Improved |
| Add HGS to EGR and HGR (ablation) | LangGoalIRL · R2R seen validation | success rate: 0.504 ± 0.0185 -> 0.530 ± 0.0138 | Improved |
| Add EGR to HGR and HGS (ablation) | LangGoalIRL · R2R seen validation | success rate: 0.503 ± 0.0117 -> 0.530 ± 0.0138 | Improved |
| Add HER to HGR without EGR or HGS (ablation) | LangGoalIRL · R2R seen validation | success rate: 0.427 ± 0.0048 -> 0.101 ± 0.0790 | Hurt |
| Add HER to the full EGR, HGR, and HGS method (ablation) | LangGoalIRL · R2R seen validation | success rate: 0.530 ± 0.0138 -> 0.521 ± 0.0093 | Hurt |
| Collect new-environment demonstrations with the goal-agnostic policy instead of random exploration (ablation) | LangGoalIRL_SS · R2R unseen validation, 15,000 demonstrations | success rate: 0.347 ± 0.0098 -> 0.491 ± 0.0065 | Improved |
Closest papers by meaning
- RLZero: Direct Policy Inference from Language Without In-Domain Supervision (NeurIPS 2025 · similarity 0.88)
- Text-Aware Diffusion for Policy Learning (NeurIPS 2024 · similarity 0.88)
- Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 2025 · similarity 0.88)
- GoalLadder: Incremental Goal Discovery with Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- The Synergy of LLMs & RL Unlocks Offline Learning of Generalizable Language-Conditioned Policies with Low-fidelity Data (ICML 2025 · similarity 0.88)
- Non-Adversarial Inverse Reinforcement Learning via Successor Feature Matching (ICLR 2025 · 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.
