Research map / Reinforcement learning / Imitation learning
Visual Adversarial Imitation Learning using Variational Models
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 |
|---|---|---|---|
| Use V-MAIL's learned model for target-task imitation without additional environment samples instead of behavior cloning | Walker Walk transfer · Walker Walk | percent of expert return: 26.8% -> 92.7% | Helped |
| Use V-MAIL's learned model for target-task imitation without additional environment samples instead of offline DAC | Walker Walk transfer · Walker Walk | percent of expert return: 8.8% -> 92.7% | Helped |
| Use V-MAIL's learned model for target-task imitation without additional environment samples instead of behavior cloning | Claw Rotate transfer · Claw Rotate | percent of expert return: 8.3% -> 97.9% | Helped |
| Use V-MAIL's learned model for target-task imitation without additional environment samples instead of offline DAC | Claw Rotate transfer · Claw Rotate | percent of expert return: -0.7% -> 97.9% | Helped |
| Use V-MAIL's learned model for target-task imitation without additional environment samples instead of direct policy transfer | Walker Walk transfer · Walker Walk | percent of expert return: 21.3% -> 92.7% | Helped |
| Use V-MAIL's learned model for target-task imitation without additional environment samples instead of direct policy transfer | Claw Rotate transfer · Claw Rotate | percent of expert return: 5.6% -> 97.9% | Helped |
| Use V-MAIL's learned model for target-task imitation without additional environment samples instead of behavior cloning | Claw Rotate transfer · Claw Rotate | percent of expert return | Helped |
| Reduce demonstrations from 10 to 5 or 1 (ablation) | V-MAIL · Cheetah Run and Walker Walk | performance | Mixed |
| Use the variational model for representation only, then train DAC (VDAC), instead of training V-MAIL in the learned model (ablation) | V-MAIL · locomotion tasks | sample efficiency: 30% less sample efficient | Hurt |
| Train adversarial imitation with model-generated on-policy rollouts rather than environment-only rollouts | V-MAIL · visual locomotion and manipulation tasks | sample efficiency | Helped |
| Train adversarial imitation with model-generated on-policy rollouts | V-MAIL · visual locomotion and manipulation tasks | learning stability | Helped |
Closest papers by meaning
- Variational Adversarial Kernel Learned Imitation Learning (AAAI 2020 · similarity 0.90)
- Robust Imitation of Diverse Behaviors (NeurIPS 2017 · similarity 0.90)
- SeMAIL: Eliminating Distractors in Visual Imitation via Separated Models (ICML 2023 · similarity 0.90)
- Diffusion-Reward Adversarial Imitation Learning (NeurIPS 2024 · similarity 0.89)
- Diffusion Imitation from Observation (NeurIPS 2024 · similarity 0.88)
- Robust Visual Imitation Learning with Inverse Dynamics Representations (AAAI 2024 · 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.