Research map / Reinforcement learning / Visual and representation RL
MCP: Learning Composable Hierarchical Control with Multiplicative Compositional Policies
Xue Bin Peng, Michael Chang, Grace Zhang, Pieter Abbeel, Sergey Levine
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 MCP instead of the strongest listed baseline, Finetune | Biped heading control · Heading: Biped | normalized return: 0.970 \pm 0.002 -> 0.976 \pm 0.002 | Improved |
| Use MCP instead of Finetune | Humanoid heading control · Heading: Humanoid | normalized return: 0.975 \pm 0.008 -> 0.970 \pm 0.003 | Hurt |
| Use MCP instead of Finetune | T-Rex heading control · Heading: T-Rex | normalized return: 0.953 \pm 0.004 -> 0.932 \pm 0.007 | Hurt |
| Use multiplicative MCP composition instead of the latent-space model | Biped box carrying · Carry: Biped | normalized return: 0.456 \pm 0.031 -> 0.575 \pm 0.032 | Improved |
| Use multiplicative MCP composition instead of the latent-space model | Biped ball dribbling · Dribble: Biped | normalized return: 0.768 \pm 0.012 -> 0.782 \pm 0.008 | Improved |
| Use multiplicative MCP composition instead of the latent-space model | Humanoid ball dribbling · Dribble: Humanoid | normalized return: 0.751 \pm 0.006 -> 0.805 \pm 0.006 | Improved |
| Use multiplicative MCP composition instead of the latent-space model | T-Rex ball dribbling · Dribble: T-Rex | normalized return: 0.115 \pm 0.013 -> 0.781 \pm 0.021 | Improved |
| Use MCP instead of Scratch | Ant holdout heading control · Holdout: Ant | normalized return: 0.951 \pm 0.093 -> 0.812 \pm 0.030 | Hurt |
| Increase primitive count from 4 to 8 (ablation) | MCP transfer policies · Transfer tasks | learning performance | No clear effect |
| Increase primitive count beyond 8, including to 32 (ablation) | MCP transfer policies · Transfer tasks | learning efficiency | Hurt |
| Fine-tune the intra-option policies and termination functions during transfer (ablation) | Option-Critic · Transfer tasks | performance | No clear effect |
Closest papers by meaning
- Learning transferable motor skills with hierarchical latent mixture policies (ICLR 2022 · similarity 0.89)
- Policy Compatible Skill Incremental Learning via Lazy Learning Interface (NeurIPS 2025 · similarity 0.88)
- CompILE: Compositional Imitation Learning and Execution (ICML 2019 · similarity 0.88)
- CoMic: Complementary Task Learning & Mimicry for Reusable Skills (ICML 2020 · similarity 0.88)
- Discovering Motor Programs by Recomposing Demonstrations (ICLR 2020 · similarity 0.87)
- Goal-Oriented Skill Abstraction for Offline Multi-Task Reinforcement Learning (ICML 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.
