SPIN @ CoRL 2026 — Structured Physical Intelligence

Overview

What does it take to achieve the physical intelligence of a five-year-old? Children learn to manipulate new objects and generalize skills from only a handful of interactions. Despite rapid progress in vision and language, today’s robots remain far less data-efficient and lack comparably rich training data.

Structured Physical Intelligence (SPIN) asks how physical priors and structured world representations can close this gap. Relevant approaches include Visual-Language-Action models (VLAs), world models, object-centric and compositional representations, affordances, and particle-based models. Robotics provides a direct test of whether these ideas support efficient generalization across locomotion, dexterous manipulation, hierarchical control, and few-shot learning.

The workshop brings researchers together to examine recent advances, current limitations, and open problems. The workshop will feature a live dexterous manipulation competition that puts today’s robots side by side with a child, with invited talks and a panel discussion.

Research Questions

This half-day workshop is organized around five questions spanning structure, representation, learning, and evaluation for physical agents:

Q1 — What structure enables data-efficient physical intelligence? How can developmental and cognitive theories of human physical knowledge inspire structured representations in robotics, and which of those inductive biases actually pay off in practice?
Q2 — How should physical agents represent the world? Object-centric models, compositional structure, affordances and object interactions, particle motion representations, and multimodal grounding each propose a different answer. Which representations best support perception, prediction, and action?
Q3 — What should be built in, and what should emerge from scale? How should engineered structural priors interact with self-supervised learning, joint-embedding predictive architectures, VLAs, and world models as data and model size grow?
Q4 — How can structured models generalize and adapt? What enables parameter-efficient fine-tuning, prompting, adapters, multi-task and continual learning, and transfer across tasks, environments, and robot embodiments?
Q5 — How should progress be understood and evaluated? Which benchmarks and mechanistic interpretability methods reveal whether a model truly reasons about physical interactions, dynamics, and failure?

The Child–Robot Dexterity Challenge

How do today’s robots compare with a child on everyday manipulation tasks?

Side by side: a robot gripper and a child's hands assembling matching wooden parts — Same task, two kinds of physical intelligence. Image is placeholder material pending final permissions and credits.

A child can quickly pick up, turn, stack, and insert unfamiliar objects. This competition invites robotic systems to attempt the same kinds of tasks, creating a clear and constructive benchmark for current progress in dexterous manipulation. It is open to all researchers, students, and practitioners from academia and industry.

How it works

The competition will feature 10 dexterous manipulation tasks using familiar objects. Each task will be selected so that it can be completed by a child, providing an intuitive point of comparison for today’s robotic systems.

Key dates and participation details

Final task list: TBD. Examples may include picking up a small object, untwisting a bottle cap, inserting an object into a matching opening, building towers, and rotating a face of a Rubik’s cube.
Competition launch: TBD
Registration deadline: TBD

Hardware

We will provide hardware for participants to use. Platform details will be announced. Teams and participants are also encouraged to bring their own identical hardware, subject to approval.

Evaluation

On the day of the competition, every task will be evaluated against its own clearly defined success criterion. A panel of neutral judges will assess each attempt consistently, and each system’s performance will be compared task by task with the child’s performance on the same activities. The detailed judging procedure, number of attempts, and reporting format are TBD.

The goal is to produce a transparent, constructive view of current capabilities across a varied set of everyday tasks and to identify promising directions for future research.

Speakers

All four invited speakers are committed, and will also join the panel discussion.

Mike Rabbat

AMI Labs

Speaker & panelist

Talk title to be announced

Jeannette Bohg

Stanford University

Speaker & panelist

Talk title to be announced

Yunzhu Li

Columbia University

Speaker & panelist

Talk title to be announced

Georgia Chalvatzaki

TU Darmstadt

Speaker & panelist

Talk title to be announced

Schedule

Tentative morning half-day schedule, in person.

8:30	Welcome — opening remarks, tasks and rules (10 min)
8:40	Keynote Talk 1 — 25 min + 5 min Q&A
9:10	Keynote Talk 2 — 25 min + 5 min Q&A
9:40	Challenge, Part 1 — 40 min.
10:20	Keynote Talk 3 — 25 min + 5 min Q&A
10:50	Keynote Talk 4 — 25 min + 5 min Q&A
11:20	Challenge, Part 2 and Judging — 40 min.
12:00	Results and Panel: What is Next in Structured Physical Intelligence Models? — 45 min.
12:45	Concluding remarks — 5 min
12:50	Adjourn