About the Workshop

The Challenge: Robotics remains data-limited, while robot learning moves toward ever larger models
The Question: Is data all we need, or can physics priors complement limited data for robot planning and control?
The Approach: A critical examination of the evidence for and against physics priors
The Format: Five invited talks, paper spotlights, question-driven posters, and a round-table debate

Data-driven learning is transforming robotics, but unlike language and vision, robotics remains fundamentally data-limited: collecting physical interaction data requires hardware, time, and supervision, and safety limits exploration. At the same time, robot learning is moving toward ever larger, general-purpose models, and substantial resources are being committed to pure data scaling. This raises a fundamental and timely question:

Is data all we need, or can physics priors complement limited data for robot planning and control?

The workshop focuses on learning-based robot planning and control, and asks when physics priors embedded in learned models and policies improve closed-loop behavior. Physics priors range from governing equations (e.g., Euler–Lagrange dynamics), conservation laws, and symmetries to contact models and physical-consistency properties such as passivity and actuator limits.

Rather than advocating for physics priors, the workshop critically examines their benefits and limitations over purely data-driven alternatives. It explores whether embedded physics may compensate for limited data and promote physical admissibility in safety-critical models and policies. It also examines when inaccurate assumptions become restrictive or harmful. Every speaker and paper author will be asked to address three questions:

(Q1) Embedding physics priors. Which physical assumptions does the method embed in the learned model or policy, and at which level: inputs and data generation, model architecture, or training loss? How do they affect accuracy and data efficiency, and when do inaccurate assumptions become harmful?
(Q2) From models to actions. Do model-level gains improve closed-loop planning and control? Where comparisons are feasible, how does the method perform against alternatives with weaker, stronger, or no physics priors, under comparable data and computational budgets (e.g., task success, generalization, closed-loop safety)? Otherwise, what evidence supports the claimed benefits, and what limitations remain?
(Q3) Scaling and generalization. As models and datasets scale, which roles for physics priors remain useful for reliable planning and control, and how does the accuracy of the assumed physics affect them?

The invited talks follow the same storyline: from embedding physics priors in learned models (Q1), to using them to plan and control robot actions (Q2), to their role as learning scales (Q3).

The answers to Q1–Q3, the disagreements and open questions raised at the poster session, and the round-table conclusions will be collected in a position paper, published online after the workshop. The paper will distinguish findings supported by comparative empirical evidence, context-dependent conclusions and methodological trade-offs, and open questions where the available evidence is insufficient.

The workshop builds on a recent survey paper by members of the organizing team (arXiv:2609.22319, GitHub repository).

Confirmed Speakers

Cosimo Della Santina Cosimo Della Santina
Associate Professor
TU Delft, The Netherlands
Ken Goldberg Ken Goldberg
Full Professor
UC Berkeley, USA
Ines Sorrentino Ines Sorrentino
Locomotion Control Tech Lead
Generative Bionics, Italy
Yuexin Ma Yuexin Ma
Associate Professor
ShanghaiTech University, China
Weiming Zhi Weiming Zhi
Assistant Professor
University of Sydney, Australia

Program

The half-day workshop features five invited talks from academia and industry, spotlight talks of 5 shortlisted papers, a question-driven poster session, and a one-hour round-table debate. The workshop will be held in person in Seoul, South Korea.

Time Event Notes
08:30 - 08:35 Opening Welcome and live poll (Slido): “Is data all we need?”
08:35 - 10:15 Invited Talks 1–4 20 min + 5 min Q&A for each talk
10:15 - 10:30 Paper Spotlights 5 shortlisted papers, 3 min per paper
10:30 - 11:00 Posters & Coffee Break Question-driven poster session with speaker-led clusters (Q1/Q2/Q3)
11:00 - 11:25 Invited Talk 5 20 min + 5 min Q&A
11:25 - 12:25 Round Table Scaling Robot Learning: How Much Do Physics Priors Matter for Planning and Control?
12:25 - 12:30 Closing Best-paper award

Question-Driven Poster Session

Accepted posters are grouped by their main question (Q1–Q3), and each cluster, led by one invited speaker, identifies one key disagreement and one unresolved research question. These become inputs to the round table and the position paper. The assignment of speakers to poster clusters will be announced here in advance.

Round Table: Scaling Robot Learning: How Much Do Physics Priors Matter for Planning and Control?

Moderated by one of the organizers, with the five invited speakers and one early-career author among the shortlisted papers. The one-hour debate follows Q1–Q3. Each part starts from the topics of one or two talks, and then involves all panelists:

  1. Which physics assumptions are worth embedding? (Q1) Starting from models learned from pixels: which assumptions improve accuracy and data efficiency, and when do wrong ones hurt?
  2. Can physics priors make planning and control better and safer? (Q2) Starting from motion planning and humanoid control: which priors are useful for safe closed-loop behavior, and what evidence shows it?
  3. Does scale make physics unnecessary, or move it? (Q3) Starting from data synthesis and agentic robotics: can physics be learned directly from data? Will agentic robotics make physics priors unnecessary, or will they move to the agentic level?
  4. Where should the community invest? The poster clusters present their disagreements and open questions. Each panelist then states one bet for the next five years (physics priors, data and scale, or both), challenged by Slido questions, before a final vote on the opening poll.

Attendees can post questions and objections on Slido, also anonymously. The talks and the round table will be recorded and made available online.

Call for Papers

We invite non-archival papers on learning-based robot planning and control. Planning includes path, motion, and task planning, and control ranges from whole-body to low-level and actuator control. Papers may:

  • embed physics priors in learned models, planners, or policies;
  • analyze the trade-off between physics priors and data;
  • show cases where physics does not help and data-driven learning performs better.

Controversial views, preliminary work, and negative results are explicitly welcome. All papers should address the questions Q1–Q3 above.

Submission Guidelines

  • Papers must follow the IEEE ICRA format and have 4–8 pages (excluding references).
  • Each submission must include a short Q1–Q3 statement (about 150 words) summarizing the physics priors used, the evidence for or against them, and their limits as data and models scale.
  • Papers are non-archival: they will be posted on the workshop website and will not be part of the ICRA proceedings.
  • Submission link: coming soon.

Review and Awards

Each submission will receive at least two reviews. Besides quality, originality, and relevance to Q1–Q3, reviewers will value clear evidence for or against physics priors (e.g., comparative evaluations or a discussion of limitations) and papers likely to stimulate discussion.

All accepted papers will be presented as posters. Based on the reviews, 5 shortlisted papers will be selected for 3-minute spotlight talks, and the author of one of them will join the round table. A Best Paper Award will be assigned by the program committee based on the reviews and the spotlight presentations.

Important Dates

Paper submissionMarch 15, 2027
Notification of acceptanceApril 12, 2027
Camera-readyApril 19, 2027
WorkshopMay 24–28, 2027

Organizers

Mattia Piccinini Mattia Piccinini
Humboldt Postdoctoral Fellow
Technical University of Munich, Germany
Thomas Beckers Thomas Beckers
Full Professor
Ulm University, Germany
Nadia Figueroa Nadia Figueroa
Assistant Professor
University of Pennsylvania, USA
Yuan Gao Yuan Gao
Doctoral Researcher
Technical University of Munich, Germany
Alice Plebe Alice Plebe
Assistant Professor
University of Trento, Italy
Gastone Pietro Rosati Papini Gastone Pietro Rosati Papini
Associate Professor
University of Trento, Italy
Johannes Betz Johannes Betz
Assistant Professor
Technical University of Munich, Germany

Contact: mattia.piccinini@tum.de


This workshop is supported by the IEEE RAS Technical Committee on Autonomous Ground Vehicles and Intelligent Transportation Systems, and by the Italian FIS 2 Call, Grant Assignment Decree No. 1236 adopted on 01/08/2023 by the Italian Ministry of University and Research (MUR), for the project Structured neural network framework for modeling and control of autonomous systems - Neu4mes, CUP E53C24003800001.