Humanoids2025 WS: Harmonising Robustness and Reliability

Harmonising Practical Robustness and Theoretical Reliability in Legged Locomotion: Fusion of Reinforcement Learning and Model-based Control

2 pm - 5 pm, October 2, 2025

@ COEX, Seoul, Korea, RM#209A

2025 IEEE-RAS 24th International Conference on Humanoid Robots, Workshop

Abstract

In the field of legged locomotion, reinforcement learning (RL) has demonstrated outstanding robustness, whereas model-based control has provided precise and physically informed motion capability. These independently developed approaches are now converging, as recent advances create opportunities for complementary integration of their respective strengths. This workshop aims to explore the practical and theoretical connections between these methodologies, highlighting their trade-offs, mutual benefits, and challenges in real-world deployment. The theme also leads us to discuss importance and requirements of hardware design and low-level control for modelling, learning, and sim-to-real transfer against real-world complexity.

Topics of Interest

Comparison and Contrast Between Model-based Control and Reinforcement Learning

Identify commonalities and differences between those approaches
Examine practitioners' perspectives on both sides (e.g., control engineers vs. RL researchers)
Discuss shared theoretical foundations such as Markov assumptions and Bellman's principle

Bridging the Gap Between Theoretical Guarantees and Real-world Complexity

Strategies to reduce trial-and-error in physical hardware
Benefits of common and standardized platforms for knowledge collection
Challenges in transferring mathematical control interests into practical robot deployment

Leveraging Control Theory in Reinforcement Learning

Improve learning efficiency using control-theoretic insights
Provide theoretical guarantees for learning processes and outcomes
Realize sim-to-real transfer and robustness evaluation

Integration and Complementary Use of Model-based Control and Reinforcement Learning

Practical fusion strategies like combining model predictive control (MPC) solutions with RL
Understanding why low-gain PID is often preferred in RL-based controllers
Exploring how soft bodies and mechanical compliance can absorb environmental complexity
Balancing precision and robustness in locomotion and manipulation tasks

Timetable

14:00 - 14:10 Opening Remarks

14:10 - 14:40 Talk #1: Dr. Takahiro Miki

14:40 - 15:10 Talk #2: Dr. Farbod Farshidian

15:10 - 15:20 Contributed Presentation by Mr. Jin Cheng

15:20 - 15:50 Interactive Poster Session with Coffee

15:50 - 16:20 Talk #3: Prof. Koh Hosoda

16:20 - 16:50 Talk #4: Prof. Jemin Hwangbo

16:50 - 17:00 Concluding Remarks

Speakers

Dr. Farbod Farshidian

Robotics and AI Institute

Prof. Koh Hosoda

Kyoto University

Prof. Jemin Hwangbo

KAIST

Dr. Takahiro Miki

ETH Zurich

Contributed Presentation

Jin Cheng, Dongho Kang, Gabriele Fadini, Guanya Shi, and Stelian Coros: RAMBO: RL-augmented Model-based Whole-body Control for Loco-manipulation. (Published in IEEE Robotics and Automation Letters, vol. 10, no. 9, pp. 9462-9469, 2025; DOI)

Call for Posters (now closed)

We invite researchers and developers to submit an extended abstract. We sincerely welcome contributors from academia, industries and professionals.

All extended abstracts are expected to present preliminary results, ongoing work, papers recently published in another conference or journals, future directions, and/or practices relevant to the topic of the workshop. They will be reviewed by the organisers, and decisions will be based on the quality of the contribution, originality, and relevance to the workshop.

All accepted contributors will present the content by posters during the interactive session of the workshop.

Publication Policy:

The workshop website will list the titles and authors' names of the accepted presentations.
All extended abstracts will NOT published anywhere: they will be used only for the review process.

Paper format:

1 or 2 pages, excluding the list of references.
In PDF that follows the IEEE RAS conference template.
If the paper includes contents which has already been published somewhere, please clarify the corresponding publication.

Submission Form

* If you do not have any Google account, please send it directly to <onishi at furo.org> .

Deadline: 7th September 2025 (AoE)

Acceptance Notification: 14 September 2025 (AoE)

Organizer