Continual Learning for Embodied Agents in Open-Ended Environments
Event details
| Date | 04.02.2026 |
| Hour | 14:00 › 16:00 |
| Speaker | Chengkun Li |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Wulfram Gerstner
Thesis advisor: Prof. Alexander Mathis
Co-examiner: Prof. Auke Ijspeert
Abstract
Embodied agents deployed in the real world face big worlds: open-ended, non-stationary environments whose task distributions cannot be fully captured by any fixed benchmark. Yet most reinforcement learning studies still rely on static task suites, which makes it difficult to evaluate meaningful progress when goals, skills, and environments evolve over time.
We hypothesize that progress in open-ended embodied intelligence must start with evaluation: principled ways to generate, select, and assess tasks that reveal what an agent can reliably do, what it forgets, and where it fails. This PhD, Continual Learning for Embodied Agents in Open-Ended Environments, will build on model-driven open-endedness, where foundation models propose novel tasks and produce executable environment specifications guided by notions of novelty and interestingness. We will develop evaluation protocols, metrics, and diagnostic tests for open-ended progress, including robustness under distribution shift and sensitivity to reward or success-specification errors.
Recent advances in scalable simulated control platforms, including modern MuJoCo-based suites, enable systematic long-horizon experimentation. Using these tools, we will iteratively refine the evaluation framework and, guided by its findings, progressively develop methods that improve stability and adaptation in big worlds.
Background papers
Exam president: Prof. Wulfram Gerstner
Thesis advisor: Prof. Alexander Mathis
Co-examiner: Prof. Auke Ijspeert
Abstract
Embodied agents deployed in the real world face big worlds: open-ended, non-stationary environments whose task distributions cannot be fully captured by any fixed benchmark. Yet most reinforcement learning studies still rely on static task suites, which makes it difficult to evaluate meaningful progress when goals, skills, and environments evolve over time.
We hypothesize that progress in open-ended embodied intelligence must start with evaluation: principled ways to generate, select, and assess tasks that reveal what an agent can reliably do, what it forgets, and where it fails. This PhD, Continual Learning for Embodied Agents in Open-Ended Environments, will build on model-driven open-endedness, where foundation models propose novel tasks and produce executable environment specifications guided by notions of novelty and interestingness. We will develop evaluation protocols, metrics, and diagnostic tests for open-ended progress, including robustness under distribution shift and sensitivity to reward or success-specification errors.
Recent advances in scalable simulated control platforms, including modern MuJoCo-based suites, enable systematic long-horizon experimentation. Using these tools, we will iteratively refine the evaluation framework and, guided by its findings, progressively develop methods that improve stability and adaptation in big worlds.
Background papers
- The Big World Hypothesis and its Ramifications for Artificial Intelligence
https://openreview.net/forum?id=Sv7DazuCn8 - OMNI-EPIC: Open-endedness via Models of Human Notions of Interestingness with Environments Programmed in Code
https://openreview.net/forum?id=Y1XkzMJpPd - MuJoCo Playground
https://arxiv.org/abs/2502.08844
Practical information
- General public
- Free