Research map / Multi-agent systems and games / Human-AI and goal-conditioned

ROTATE: Regret-driven Open-ended Training for Ad Hoc Teamwork

NeurIPS 2026 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Collaborative multi-agent perception · 0 citations (OpenAlex)

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.

Reformulates ad hoc teamwork as an open-ended minimax-regret game and proposes ROTATE, which generates competence-bounded per-state regret-maximizing teammates and trains an ego agent against a growing population buffer.

Problem

Ad hoc teamwork requires ego agents to cooperate with unseen teammates, but two-stage methods use fixed teammate sets with limited coverage, and adversarial teammate generation can produce incompetent (sabotaging) teammates that hinder learning.

Approach

Alternates between an ego agent and a teammate generator that maximizes the ego's cooperative regret over all visited states (not just initial states), with a competence-bounding return term (λ-weighted teammate-BR return) that provably bounds teammate competence on encountered states, plus a population buffer of all past teammates.

Key results

• ROTATE attains the highest mean normalized return on five of six tasks; Minimax Return is best on AA.
• On LBF, ROTATE achieves 89.5 unnormalized return versus 85.8 for the best baseline CoMeDi.
• In the sabotage matrix game, ROTATE teammates have near-zero sabotage probability while per-trajectory regret yields over a third of states with P(S) near 1.0.

Surprising findings

• Maximizing per-trajectory regret can yield teammates that minimize returns with every ego agent (matrix game example).
• Higher competence weights raise training returns but intermediate weights give the best evaluation returns on Cramped Room/Counter Circuit.

Limitations

• Each open-ended iteration trains teammate and best response from scratch, which is computationally expensive.
• Theory does not explain why bounding competence on teammate-ego interactions helps ego learning.
• Uniform population sampling; worst-case buffer sampling left for future work.

Experimental setup

Datasets: Overcooked (CR, AA, CC, CoR, FC layouts) Level-Based Foraging human gameplay dataset for Overcooked (human proxy)

Benchmarks: Overcooked (CR, AA, CC, CoR, FC) Level-Based Foraging

Metrics: normalized mean return IQM median probability of improvement worst-case normalized return

Baselines: PAIRED Minimax Return Fictitious Co-Play BRDiv CoMeDi COLE

Builds on: unsupervised environment design PAIRED Fictitious Co-Play BRDiv CoMeDi COLE PPO

Closest papers by meaning

• Ad Hoc Teamwork via Offline Goal-Based Decision Transformers (ICML 2025 · similarity 0.91)
• Minimum Coverage Sets for Training Robust Ad Hoc Teamwork Agents (AAAI 2024 · similarity 0.90)
• Open Ad Hoc Teamwork with Cooperative Game Theory (ICML 2024 · similarity 0.90)
• N-agent Ad Hoc Teamwork (NeurIPS 2024 · similarity 0.89)
• Adaptively Coordinating with Novel Partners via Learned Latent Strategies (NeurIPS 2025 · similarity 0.88)
• LLM-Assisted Semantically Diverse Teammate Generation for Efficient Multi-agent Coordination (ICML 2025 · similarity 0.87)