Research map / Multi-agent systems and games / Human-AI and goal-conditioned
OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code
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.
Introduces OMNI-EPIC, which uses foundation models to generate learnable and interesting tasks as Python environment and reward code in PyBullet, training DreamerV3 agents on them.
Problem
Automatically generating open-ended learning environments is limited to narrow, manually predefined distributions; the paper targets generating any learnable and interesting simulatable task.
Approach
Extends OMNI with environment programming: LLMs propose next interesting task descriptions from a growing archive, generate PyBullet environment and reward code, apply a post-generation model of interestingness, and train RL agents, using LLM-written success-checking code as the success detector.
Key results
- OMNI-EPIC achieves significantly higher cell coverage of archive diversity than both controls (p < 0.05, Mann-Whitney U test) in long simulated-learning runs.
- The success detector agreed with 50 human participants' assessments at a 72.7% alignment rate.
- Transfer learning from archive policies raised median success rate to 63.2% versus 36.8% when training from scratch (2M steps per task).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add transfer learning between tasks instead of training each task from scratch (ablation) | OMNI-EPIC agents · 25 tasks, 5 replications | success rate: 36.8\% -> 63.2\% | Helped |
| sample the most similar archive tasks instead of sampling in-context examples uniformly (ablation) | OMNI-EPIC task generator · Short run with learning | tasks successfully learned | Helped |
| include failed task examples in the task-generator context (ablation) | OMNI-EPIC task generator · Short run with learning | tasks successfully learned | Helped |
| include learning-progress notions in the task-generator prompt (ablation) | OMNI-EPIC task generator · Short run with learning | tasks successfully learned | Helped |
| use a task archive and models of interestingness | OMNI-EPIC · Long runs with simulated learning | cell coverage | Helped |
| use a task archive and models of interestingness | OMNI-EPIC · Short runs with learning | cell coverage | Mixed |
| generate tasks with an archive and models of interestingness rather than either control | OMNI-EPIC · Short runs with RL training | ANNECS-OMNI | Helped |
Surprising findings
- Without a notion of learning progress in the task generator prompt, no tasks were successfully learned in the ablation run.
- Current VLMs were not accurate enough as success detectors, so LLM-generated success-checking code was used instead.
Limitations
- Not yet Darwin Complete: restricted to the PyBullet simulator.
- Short runs limited by academic compute; RL results are from only 5 short runs.
- Success detection relies on LLM-generated code since current VLMs were not accurate enough.
Experimental setup
Datasets: task description seeds (3 per run)
Benchmarks: OMNI-EPIC generated task archives (cell coverage) ANNECS-OMNI
Metrics: cell coverage of archive diversity ANNECS-OMNI ANNECS success rate agreement with human evaluation
Baselines: OMNI-EPIC w/o archive OMNI-EPIC Learning Progress Only
Builds on: OMNI DreamerV3 Gymnasium ANNECS PyBullet
Closest papers by meaning
- OMNI: Open-endedness via Models of human Notions of Interestingness (ICLR 2024 · similarity 0.91)
- Agent-Omit: Adaptive Context Omission for Efficient LLM Agents (ICML 2026 · similarity 0.85)
- Intelligent Go-Explore: Standing on the Shoulders of Giant Foundation Models (ICLR 2025 · similarity 0.85)
- AgentOdyssey: Open-Ended Long-Horizon Text Game Generation for Test-Time Continual Learning Agents (NeurIPS 2026 · similarity 0.85)
- A Single Goal is All You Need: Skills and Exploration Emerge from Contrastive RL without Rewards, Demonstrations, or Subgoals (ICLR 2025 · similarity 0.85)
- PLATO: Pointer Learner for Agent and Task Openness (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.