Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Compositional Plan Vectors
Coline Devin, Daniel Geng, Pieter Abbeel, Trevor Darrell, Sergey Levine
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 |
|---|---|---|---|
| remove embedding subtraction input (TE-FULL instead of CPV-FULL) (ablation) | CPV-FULL · Crafting composition 1+1 | success rate: 76 ± 3 -> 60 ± 1 | Hurt |
| remove embedding subtraction input (TE-FULL instead of CPV-FULL) (ablation) | CPV-FULL · Crafting 4 skills | success rate: 73 ± 2 -> 61 ± 8 | Hurt |
| remove both regularizations (CPV-NAIVE vs CPV-FULL) (ablation) | CPV-FULL · Crafting 4 skills | success rate: 73 ± 2 -> 51 ± 8 | Hurt |
| add pair loss only (ablation) | CPV · Crafting 8 skills | success rate: 19 ± 5 -> 44 ± 14 | Improved |
| add pair loss only (ablation) | CPV · Crafting composition 1+1 | success rate: 31 ± 16 -> 2 ± 3 | Hurt |
| add homomorphism loss (ablation) | CPV · Crafting composition 1+1 | success rate: 31 ± 16 -> 71 ± 8 | Improved |
| add both pair and homomorphism losses to homomorphism-only (ablation) | CPV-HOM · Crafting 4 skills | success rate: 63 ± 3 -> 73 ± 2 | Improved |
| CPV-FULL instead of TECNet | TECNet · Crafting composition 1+1 | success rate: 59 ± 11 -> 76 ± 3 | Improved |
| CPV-FULL instead of naive model | Naive · Crafting 16 skills | success rate: 7 ± 2 -> 28 ± 6 | Improved |
| CPV-FULL instead of TECNet | TECNet · Pick and place composition 1,1 | success rate: 33 ± 4 -> 56 ± 6 | Improved |
| remove subtraction (TE-PLAIN vs CPV-PLAIN) (ablation) | CPV-PLAIN · Pick and place composition 1,1 | success rate: 52 ± 2 -> 22 ± 2 | Hurt |
| add pair loss only (ablation) | CPV-PLAIN · Pick and place composition 1,1 | success rate: 52 ± 2 -> 7 ± 1 | Hurt |
Closest papers by meaning
- Robust Subtask Learning for Compositional Generalization (ICML 2023 · similarity 0.89)
- Temporal Representation Alignment: Successor Features Enable Emergent Compositionality in Robot Instruction Following (NeurIPS 2025 · similarity 0.89)
- Synthesized Policies for Transfer and Adaptation across Tasks and Environments (NeurIPS 2018 · similarity 0.88)
- CompILE: Compositional Imitation Learning and Execution (ICML 2019 · similarity 0.88)
- Compositional Automata Embeddings for Goal-Conditioned Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
- RoboDreamer: Learning Compositional World Models for Robot Imagination (ICML 2024 · similarity 0.87)
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.
