Research map / Reinforcement learning / Imitation learning
An Imitation from Observation Approach to Transfer Learning with Dynamics Mismatch
Siddharth Desai, Ishan Durugkar, Haresh Karnan, Garrett Warnell, Josiah Hanna, Peter Stone
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 adversarial imitation-from-observation action transformation instead of GAT | Transfer from source to target with dynamics mismatch · Minitaur target environment | target performance relative to optimal: more than 80% vs at most 50% | Improved |
| Use GARAT rather than GAT for source-environment grounding | InvertedPendulumHeavy · InvertedPendulum domain | per-step transition error | Improved |
| Use GARAT rather than GAT for transfer | Transfer methods under dynamics mismatch · MuJoCo domains except HopperHighFriction | target-environment return | Improved |
| Use GARAT rather than related transfer methods | Transfer methods under dynamics mismatch · HopperHighFriction | target-environment return | No clear effect |
| Use GARAT with a few thousand target transitions rather than training directly in the target environment | Policy transfer · Most evaluated domains | target-environment return: 1 million transitions | No clear effect |
| Use GARAT rather than GAT for grounding with limited target data | Action transformation · InvertedPendulum domain | per-step transition error: a single trajectory vs 100 trajectories | Improved |
| Use GARAT rather than related methods for transfer | Policy transfer · Ant in PyBullet | target-environment return | Improved |
| Use GARAT rather than related methods for transfer | Policy transfer · Walker | target-environment return | Improved |
| Use alternate imitation-from-observation backend algorithms instead of the selected backend (ablation) | GARAT · Sim-to-real performance | target-environment return | No clear effect |
| Use other divergence measures instead of the losses detailed in the algorithm (ablation) | GARAT · Transfer experiments | target-environment return | No clear effect |
Closest papers by meaning
- Off-Dynamics Reinforcement Learning via Domain Adaptation and Reward Augmented Imitation (NeurIPS 2024 · similarity 0.90)
- Diffusion Imitation from Observation (NeurIPS 2024 · similarity 0.89)
- Generative Adversarial Imitation Learning (NeurIPS 2016 · similarity 0.89)
- Latent Wasserstein Adversarial Imitation Learning (ICLR 2026 · similarity 0.89)
- Learning from Demonstration: Provably Efficient Adversarial Policy Imitation with Linear Function Approximation (ICML 2022 · similarity 0.89)
- EvIL: Evolution Strategies for Generalisable Imitation Learning (ICML 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.
