Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Hierarchically Decoupled Imitation For Morphological Transfer
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 |
|---|---|---|---|
| Align the target low-level with the source low-level using discriminative imitation (ablation) | PointMass→Ant transfer · Waypoint | return: 483 \pm 39 -> 546\pm33 | Helped |
| Align the target low-level with the source low-level using discriminative imitation (ablation) | PointMass→Ant transfer · Maze End | success rate: .30 \pm .08 -> .55 \pm .13 | Helped |
| Align the target low-level with the source low-level using discriminative imitation (ablation) | PointMass→Quadruped transfer · Maze Sampled | success rate: .78 \pm .08 -> .89 \pm .06 | Helped |
| Align the target low-level with the source low-level using discriminative imitation (ablation) | 3Link→4Link transfer · Peg Insert | success rate: $.80 \pm .18$ -> $.93 \pm 0.05$ | Helped |
| Align the target low-level with the source low-level using discriminative imitation (ablation) | PointMass→3Link transfer · Block Push 1 | success rate: $.17 \pm .08$ -> $.34 \pm .10$ | Helped |
| Align the target low-level with the source low-level using discriminative imitation (ablation) | 2Link→3Link transfer · Block Push 2 | success rate: $.61 \pm .16$ -> $.42 \pm .11$ | Hurt |
| Use online rather than offline transition data to train the discriminator (ablation) | PointMass→Ant transfer · Maze End | success rate: 0.72\pm0.03 -> 0.63\pm0.04 | Hurt |
| Use online rather than offline transition data to train the discriminator (ablation) | PointMass→Ant transfer · Waypoint | return: 546.06\pm14.78 -> 467.22\pm20.61 | Hurt |
| Finetune the transferred high-level policy (ablation) | PointMass→Ant · Step maze | success rate: $.32 \pm .16$ -> $.53 \pm .09$ | Helped |
| Add KL regularization while finetuning the transferred high-level policy (ablation) | PointMass→Ant · Step maze | success rate: $.53 \pm .09$ -> $.52 \pm .13$ | No clear effect |
| Train the Ant high-level from scratch instead of transferring from PointMass | Ant · Step maze | success rate: $.53 \pm .09$ -> $.59 \pm .05$ | Helped |
Closest papers by meaning
- Knowledge Diversion for Efficient Morphology Control and Policy Transfer (ICML 2026 · similarity 0.88)
- Hierarchical Equivariant Policy via Frame Transfer (ICML 2025 · similarity 0.88)
- Hierarchical Few-Shot Imitation with Skill Transition Models (ICLR 2022 · similarity 0.88)
- Distilling Morphology-Conditioned Hypernetworks for Efficient Universal Morphology Control (ICML 2024 · similarity 0.88)
- GO-DICE: Goal-Conditioned Option-Aware Offline Imitation Learning via Stationary Distribution Correction Estimation (AAAI 2024 · similarity 0.87)
- Policy Decorator: Model-Agnostic Online Refinement for Large Policy Model (ICLR 2025 · similarity 0.86)
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.