Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Fighting Copycat Agents in Behavioral Cloning from Observation Histories
Chuan Wen, Jierui Lin, Trevor Darrell, Dinesh Jayaraman, 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 |
|---|---|---|---|
| use observation histories instead of a single observation | behavioral cloning · MuJoCo partially observed tasks | cumulative reward | Mixed |
| add target-conditioned adversarial prediction of the previous action | behavioral cloning with observation histories · MuJoCo partially observed tasks | cumulative reward: 592 -> 1296 | Improved |
| add target-conditioned adversarial prediction of the previous action | behavioral cloning with observation histories · six MuJoCo environments | action-prediction MSE given past actions: 0.66±0.04 -> 2.20 ± 0.06 (Ant); 1.07±0.16 -> 2.26 ± 0.12 (Hopper); 0.18±0.01 -> 1.33 ± 0.02 (Humanoid); 0.32±0.05 -> 2.9 | Improved |
| add an information bottleneck without an adversary (ablation) | behavioral cloning with observation histories · MuJoCo partially observed tasks | cumulative reward | Improved |
| add target-conditioning to the adversary (ablation) | adversarial behavioral cloning · MuJoCo partially observed tasks | cumulative reward | Improved |
| add target-conditioning to the adversary (ablation) | adversarial behavioral cloning · Hopper | next-action prediction test MSE: 14.62 ± 1.41 -> 9.48 ± 1.89 | Improved |
| add an information bottleneck to TCA (ablation) | adversarial behavioral cloning · MuJoCo partially observed tasks | cumulative reward: Ant: 1970 ± 107 -> 2150 ±34; Hopper: 683 ± 132 -> 1086 ±262; Humanoid: 696 ±48 -> 671 ±61; Reacher: -57 ± 4 -> -54 ±4; W | Mixed |
| remove target-conditioning from the adversary (ablation) | adversarial behavioral cloning · Hopper | previous-action prediction test MSE: 5.97 ± 1.32 | Hurt |
| add a target-conditioned adversary and information bottleneck | BC-OH · MuJoCo partially observed tasks | cumulative reward: 592 -> 1296 (Walker2d) | Mixed |
| use Dropout-BC on historical observations | behavioral cloning with observation histories · MuJoCo partially observed tasks | cumulative reward | Hurt |
| increase DAGGER expert queries from 100 to 1k | DAGGER · MuJoCo partially observed tasks | cumulative reward: 2090 -> 2240 (Ant); 978 -> 1120 (Hopper); 688* -> 812* (Humanoid); -52 -> -15 (Reacher); 701 -> 2170 (Walker2d); 1080 -> | Improved |
| use observation histories rather than a single observation | behavioral cloning · MuJoCo partially observed tasks | held-out demonstration test loss: 0.1079 -> 0.0829 (Ant); 0.0077 -> 0.0049 (Hopper); 0.5710 -> 0.5470 (Humanoid); 0.0008 -> 0.0007 (Reacher); 0.0410 -> 0. | Improved |
Closest papers by meaning
- Imitation by Predicting Observations (ICML 2021 · similarity 0.91)
- Regularized Behavior Cloning for Blocking the Leakage of Past Action Information (NeurIPS 2023 · similarity 0.90)
- Resolving Copycat Problems in Visual Imitation Learning via Residual Action Prediction (ECCV 2022 · similarity 0.90)
- Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning (NeurIPS 2024 · similarity 0.89)
- Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning (NeurIPS 2025 · similarity 0.88)
- Diffusion Imitation from Observation (NeurIPS 2024 · 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.
