Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Ad Hoc Teamwork via Offline Goal-Based Decision Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Extends ad hoc teamwork to the offline setting and proposes TAGET, a hierarchical goal-conditioned Decision Transformer that predicts teammate-aware return-to-go and sub-goals to act with unseen teammates from a fixed multi-agent dataset.
Problem
Ad hoc teamwork requires an ego agent to collaborate with unknown teammates, but existing AHT methods need online interaction and carefully designed teammate pools, which are infeasible in many real settings. The paper targets learning such a policy purely from offline multi-agent interaction data u
Approach
A hierarchical sequence model in which a high-level module predicts a teammate-aware return-to-go (TA-RTG) and a teammate-aware sub-goal (TA-Goal, a predicted future global state) from local observations via two regularized encoders, and a low-level goal-conditioned Decision Transformer generates actions; trajectory mirroring rotates the ego role across agents to enlarge the dataset.
Key results
- TAGET outperforms DT, Prompt-DT, ODITS-off, LIAM-off and other baselines across PP, LBF and Overcooked with unseen teammate sets, reported as an average improvement of 37.83%.
- Ablation shows removing the TA-Goal decoder causes the largest performance drop, especially in Overcooked.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TAGET instead of the evaluated baselines | Offline ad hoc teamwork · Predator-prey, Level-Based Foraging, and Overcooked | return: 37.83% | Helped |
| Remove the TA-Goal decoder (ablation) | TAGET · Overcooked | return | Hurt |
| Remove offline data pre-processing with trajectory mirroring (ablation) | TAGET · Predator-prey, Level-Based Foraging, and Overcooked | return | Hurt |
Limitations
- Evaluated only on three simple grid-world style MARL environments.
- Relies on a pre-collected offline dataset whose teammate coverage may be limited.
- Reported gains are averaged over environments; variability of baselines such as ODITS and LIAM is high.
Experimental setup
Datasets: Predator-prey offline dataset Level-Based Foraging offline dataset Overcooked offline dataset
Benchmarks: Predator-prey (PP-4, PP-8) Level-Based Foraging (LBF-4, LBF-8) Overcooked (Overcooked-4, Overcooked-8)
Metrics: average episodic return 95% confidence interval
Baselines: Decision Transformer Prompt-DT ODITS-off LIAM-off CQL MADT
Builds on: Decision Transformer ODITS LIAM Prompt-DT CSP
Closest papers by meaning
- ROTATE: Regret-driven Open-ended Training for Ad Hoc Teamwork (NeurIPS 2026 · similarity 0.91)
- Open Ad Hoc Teamwork with Cooperative Game Theory (ICML 2024 · similarity 0.89)
- N-agent Ad Hoc Teamwork (NeurIPS 2024 · similarity 0.88)
- Minimum Coverage Sets for Training Robust Ad Hoc Teamwork Agents (AAAI 2024 · similarity 0.88)
- LLM-Assisted Semantically Diverse Teammate Generation for Efficient Multi-agent Coordination (ICML 2025 · similarity 0.86)
- NestRL: A Nested Training Regime for Mutual Adaptation in Human–AI Teaming (NeurIPS 2026 · similarity 0.85)
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.