Research map / Reinforcement learning / Imitation learning
Imitation with Neural Density Models
Kuno Kim, Akshat Jindal, Yang Song, Jiaming Song, Yanan Sui, Stefano Ermon
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 NDI+EBM rather than the strongest listed baseline | Imitation learning with one demonstration · Hopper | task reward: 3261 \pm 533 -> 3458 \pm 210 | Improved |
| use NDI+EBM rather than the strongest listed baseline | Imitation learning with one demonstration · Half-Cheetah | task reward: 3456 \pm 401 -> 4511 \pm 569 | Improved |
| use NDI+EBM rather than the strongest listed baseline | Imitation learning with one demonstration · Walker | task reward: 3957 \pm 253 -> 5061 \pm 135 | Improved |
| use NDI+EBM rather than the strongest listed baseline | Imitation learning with one demonstration · Ant | task reward: 2299 \pm 519 -> 4293 \pm 431 | Improved |
| use NDI+MADE rather than the strongest listed baseline | Imitation learning with one demonstration · Humanoid | task reward: 364 \pm 50 -> 6088 \pm 689 | Improved |
| use NDI+MADE density estimation instead of NDI+EBM | NDI · Ant, one demonstration | task reward: 4293 \pm 431 -> 555 \pm 311 | Hurt |
| set lambda_f to 0.0001 rather than 0.005 (ablation) | NDI+EBM · Mujoco tasks | task reward: Hopper: 3458 \pm 210 -> 3506 \pm 188; Half-Cheetah: 4511 \pm 569 -> 5697 \pm 805; Walker: 5061 \pm 135 -> 5171 \pm 157; | Mixed |
| set lambda_f to 0.0001 rather than 0.005 (ablation) | NDI+EBM · Mujoco tasks | imitation KL (lower is better): Hopper: 0.11 \pm 0.02 -> 0.15 \pm 0.05; Half-Cheetah: 0.17 \pm 0.09 -> 0.32 \pm 0.15; Walker: 0.22 \pm 0.14 -> 0.25 \pm | Hurt |
| set lambda_f to 0.1 rather than 0.005 (ablation) | NDI+EBM · Mujoco tasks | task reward: Hopper: 3458 \pm 210 -> 1057 \pm 29; Half-Cheetah: 4511 \pm 569 -> 103 \pm 59; Walker: 5061 \pm 135 -> 2710 \pm 501; Ant | Hurt |
| set lambda_f to 0.1 rather than 0.005 (ablation) | NDI+EBM · Mujoco tasks | imitation KL (lower is better): Hopper: 0.11 \pm 0.02 -> 0.78 \pm 0.13; Half-Cheetah: 0.17 \pm 0.09 -> 1.41 \pm 0.51; Walker: 0.22 \pm 0.14 -> 0.41 \pm | Hurt |
| use lambda_f=0.005 rather than lambda_f=0.0001 (ablation) | NDI+EBM · Mujoco tasks | imitation KL (lower is better): Hopper: 0.15 \pm 0.05 -> 0.11 \pm 0.02; Half-Cheetah: 0.32 \pm 0.15 -> 0.17 \pm 0.09; Walker: 0.25 \pm 0.04 -> 0.22 \pm | Improved |
| add BC pretraining | NDI+MADE · Ant, one demonstration | task reward: 555 \pm 311 -> 4014 \pm 105 | Improved |
Closest papers by meaning
- Imitation Learning by Reinforcement Learning (ICLR 2022 · similarity 0.88)
- Imitation by Predicting Observations (ICML 2021 · similarity 0.88)
- Imitation Learning from Vague Feedback (NeurIPS 2023 · similarity 0.88)
- Exploring Gradient Explosion in Generative Adversarial Imitation Learning: A Probabilistic Perspective (AAAI 2024 · similarity 0.86)
- Reward-free World Models for Online Imitation Learning (ICML 2025 · similarity 0.86)
- Hybrid Inverse Reinforcement Learning (ICML 2024 · similarity 0.86)
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.
