Research map / Reinforcement learning / Imitation learning
Robust Imitation of a Few Demonstrations with a Backwards Model
Jung Yeon Park, Lawson L.S. Wong
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 |
|---|---|---|---|
| train the policy on backward-model traces as well as demonstrations | BC · Fetch Push | robustness (%): 12.1 ± 0.3 -> 14.6 ± 0.6 | Improved |
| train the policy on backward-model traces as well as demonstrations | BC · Fetch PickAndPlace | robustness (%): 4.1 ± 0.1 -> 17.5 ± 0.9 | Improved |
| train the policy on backward-model traces as well as demonstrations | BC · Point mazes | robustness (%): 49.0 ± 1.9 -> 47.8 ± 3.5; 36.8 ± 3.4 -> 38.6 ± 3.4; 33.7 ± 1.5 -> 38.9 ± 2.3 | Mixed |
| train the policy on backward-model traces as well as demonstrations | BC · Ant mazes | robustness (%): 63.0 ± 1.0 -> 64.8 ± 1.5; 33.2 ± 0.9 -> 29.1 ± 0.8; 21.7 ± 0.6 -> 17.6 ± 0.5 | Mixed |
| train the policy on backward-model traces as well as demonstrations | BC · Adroit Relocate | robustness (%): 7.9 ± 0.7 -> 13.3 ± 1.0 | Improved |
| replace the backwards dynamics model with a forward dynamics model (ablation) | BMIL · Fetch Push | robustness (%): 12.1 ± 0.3 -> 12.4 ± 0.6 | No clear effect |
| replace the backwards dynamics model with a forward dynamics model (ablation) | BMIL · Fetch PickAndPlace | robustness (%): 4.1 ± 0.1 -> 4.1 ± 0.2 | No clear effect |
| perturb the first generated action during trace generation (ablation) | BMIL · Fetch PickAndPlace | robustness (%) | Improved |
| increase the number of demonstrations (ablation) | BC and BMIL · Fetch PickAndPlace | robustness (%) | Improved |
| increase BC gradient steps to match or exceed BMIL compute (ablation) | BC · Fetch domains | robustness (%) | Hurt |
| train the backwards model before training the policy (ablation) | BMIL · Fetch Push | robustness (%): 14.6 ± 0.6 -> 13.2 ± 1.0 | No clear effect |
| train the backwards model before training the policy (ablation) | BMIL · Fetch PickAndPlace | robustness (%): 17.5 ± 0.9 -> 19.9 ± 2.4 | No clear effect |
Closest papers by meaning
- Learning When to Stop: Selective Imitation Learning Under Arbitrary Dynamics Shift (NeurIPS 2026 · similarity 0.90)
- Uncertainty-Based Smooth Policy Regularisation for Reinforcement Learning with Few Demonstrations (NeurIPS 2025 · similarity 0.90)
- How to Leverage Diverse Demonstrations in Offline Imitation Learning (ICML 2024 · similarity 0.90)
- Imitation Learning from Purified Demonstrations (ICML 2024 · similarity 0.90)
- Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning (NeurIPS 2024 · similarity 0.90)
- Robust Visual Imitation Learning with Inverse Dynamics Representations (AAAI 2024 · similarity 0.89)
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.
