Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Unsupervised Partner Design Enables Robust Ad-hoc Teamwork
Constantin Ruhdorfer, Matteo Bortoletto, Victor Oei, Anna Penzkofer, Andreas Bulling
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 adaptive partner generation with learnability selection (UPD) instead of E3T with fixed epsilon | E3T · Overcooked-AI (5 layouts avg) | average return: 78.8 -> 94.4 | Improved |
| Use UPD instead of E3T across swept epsilon values | E3T · Level-Based Foraging, ten evaluation partners | average episodic return | Improved |
| Add persistent action bias mask to partner generation (ablation) | UPD · Overcooked-AI (5 layouts avg) | average return: 87.9 -> 94.4 | Improved |
| Add learnability-based partner selection (ablation) | UPD · Overcooked-AI (5 layouts avg) | average return: 90.8 -> 94.4 | Improved |
| Add learnability selection on Cramped Room layout (ablation) | UPD · Overcooked-AI Cramped Room | average return: 110.8 -> 107.5 | No clear effect |
| Sample epsilon from U(0,1) with random partner generation only (no learnability) (ablation) | E3T · Overcooked-AI (5 layouts avg) | average return: 78.8 -> 90.8 | Improved |
| Use joint partner-environment curriculum (JUPD) instead of SFL levels with E3T partners | SFLE3T · Overcooked Generalisation Challenge 5x5 | average return: 44.0 -> 58.9 | Improved |
| Use self-play on random levels (CEC) instead of joint curriculum | JUPD · Overcooked Generalisation Challenge 5x5 | average return: 58.9 -> 23.9 | Hurt |
| Use UPD instead of population-based MEP and FCP | FCP and MEP populations of 48 · Overcooked-AI (5 layouts avg) | average return: 75.3 -> 94.4 | Improved |
| Train ego with UPD instead of SP, MEP, E3T with human partners | Human-AI Overcooked-AI · Human study, 12 participants, AA/CR/CC | return and survey ratings | Improved |
Closest papers by meaning
- Minimum Coverage Sets for Training Robust Ad Hoc Teamwork Agents (AAAI 2024 · similarity 0.89)
- Diversity Is Not All You Need: Training A Robust Cooperative Agent Needs Specialist Partners (NeurIPS 2024 · similarity 0.89)
- Adaptively Coordinating with Novel Partners via Learned Latent Strategies (NeurIPS 2025 · similarity 0.88)
- Cross-environment Cooperation Enables Zero-shot Multi-agent Coordination (ICML 2025 · similarity 0.88)
- ROTATE: Regret-driven Open-ended Training for Ad Hoc Teamwork (NeurIPS 2026 · similarity 0.88)
- Partner Modelling Emerges in Recurrent Agents (But Only When It Matters) (NeurIPS 2025 · 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.
