Research map / Reinforcement learning / Imitation learning
Resolving Copycat Problems in Visual Imitation Learning via Residual Action Prediction
Chia-Chi Chuang, Donglin Yang, Chuan Wen, Yang Gao
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 |
|---|---|---|---|
| add action residual prediction memory module with policy stream | BCOH (CILRS) · CARLA NoCrash Dense (train town & weather) | #SUCCESS: 34.1 -> 52.0 | Improved |
| add stop-gradient on memory feature (ablation) | OURS · CARLA NoCrash Dense (train town & weather) | #SUCCESS: 45.0 -> 52.0 | Improved |
| train memory module on residual a_t - a_{t-1} instead of a_t (ablation) | OURS · CARLA NoCrash Dense (train town & weather) | #SUCCESS: 41.3 -> 52.0 | Improved |
| train memory module on residual instead of a_{t-1} (ablation) | OURS · CARLA NoCrash Dense (train town & weather) | #SUCCESS: 47.0 -> 52.0 | Improved |
| add policy stream with current observation to memory module (ablation) | OURS · CARLA NoCrash Dense (train town & weather) | #SUCCESS: 0.0 -> 52.0 | Improved |
| add residual prediction branches a_t - a_{t-i} (2 branches) (ablation) | OURS · CARLA NoCrash Dense (train town & weather) | #SUCCESS: 52.0 -> 52.6 | No clear effect |
| use 4 residual branches (ablation) | OURS · CARLA NoCrash Dense (train town & weather) | #SUCCESS: 52.0 -> 50.7 | Hurt |
| use two-stream architecture without residual objective (ablation) | BCOH · CARLA NoCrash Dense (train town & weather) | #SUCCESS: 34.1 -> 23.7 | Hurt |
| use two-stream architecture without residual objective (ablation) | Keyframe · CARLA NoCrash Dense (train town & weather) | #SUCCESS: 41.9 -> 38.7 | Hurt |
| use residual memory method instead of BCOH on intervention test | BCOH · CARLA validation set | behavior change %: 40.88 -> 16.18 | Improved |
| use residual memory feature, measured by previous-action regression MSE | BCOH · CARLA validation set | MSE loss x10^-2: 4.83 -> 7.50 | Improved |
| use residual memory method instead of BCOH on green light passing | BCOH · CARLA NoCrash Dense | green light passing %: 30.6 -> 66.3 | Improved |
Closest papers by meaning
- Fighting Copycat Agents in Behavioral Cloning from Observation Histories (NeurIPS 2020 · similarity 0.90)
- InfoGAIL: Interpretable Imitation Learning from Visual Demonstrations (NeurIPS 2017 · similarity 0.90)
- Robust Visual Imitation Learning with Inverse Dynamics Representations (AAAI 2024 · similarity 0.90)
- Videos are Sample-Efficient Supervisions: Behavior Cloning from Videos via Latent Representations (NeurIPS 2025 · similarity 0.89)
- Diffusion Imitation from Observation (NeurIPS 2024 · similarity 0.89)
- Latent Diffusion Planning for Imitation Learning (ICML 2025 · similarity 0.88)
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.
