Research map / Multi-agent systems and games / Human-AI and goal-conditioned
A Domain-Independent Agent Architecture for Adaptive Operation in Evolving Open Worlds
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.
Presents HYDRA, a domain-independent model-based agent architecture that detects, characterizes, and accommodates novelties in mixed discrete-continuous environments by monitoring PDDL+ plan inconsistency and reward divergence and repairing the PDDL+ model via heuristic search.
Problem
Model-based planning agents fail when environment dynamics evolve away from their domain models in open worlds, and existing model-repair and replanning methods do not handle mixed discrete-continuous physics-based domains.
Approach
A novelty meta-reasoning layer around a PDDL+ planning agent: monitors compare expected plan trajectories (simulated under the PDDL+ model) with observed trajectories and reward predictions; when inconsistency exceeds a threshold, a heuristic search over model manipulation operators repairs the domain model to fit the novel environment.
Key results
- Reports that PDDL+ model repair via heuristic search can restore agent performance in CartPole++ and ScienceBirds after novelty is introduced (no specific numbers in the provided text).
- Shows that the novelty monitors (plan inconsistency, reward divergence, unknown objects) can detect novelties in three domains.
- Notes that existing PDDL+ planners UPMurphi, DiNo, and ENHSP did not scale to the development domains, motivating use of Nyx.
Surprising findings
- Off-the-shelf RL policies trained before novelty can become ineffective when the environment dynamics change, as the authors argue empirically.
Limitations
- Limited to single persistent novelty per trial and between-episode novelty introduction.
- Inconsistency thresholds are set manually by the model designer per domain.
- Repairs are restricted to a designated set of repairable fluents and predefined model manipulation operators.
Experimental setup
Datasets: CartPole++ ScienceBirds PogoStick (Polycraft)
Benchmarks: CartPole++ ScienceBirds PogoStick
Metrics: inconsistency score reward divergence task performance under novelty
Baselines: UPMurphi DiNo ENHSP PPO DQN
Builds on: PDDL+ Nyx Belief-Desire-Intention General Diagnosis Engine
Closest papers by meaning
- Agent2World: Learning to Generate Symbolic World Models via Adaptive Multi-Agent Feedback (NeurIPS 2026 · similarity 0.87)
- WorldCoder, a Model-Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment (NeurIPS 2024 · similarity 0.86)
- Craftium: Bridging Flexibility and Efficiency for Rich 3D Single- and Multi-Agent Environments (ICML 2025 · similarity 0.86)
- VAGEN: Reinforcing World Model Reasoning for Multi-Turn VLM Agents (NeurIPS 2025 · similarity 0.86)
- ADAM: An Embodied Causal Agent in Open-World Environments (ICLR 2025 · similarity 0.86)
- Training One Model to Master Cross-Level Agentic Actions via Reinforcement Learning (CVPR 2026 · 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.