Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction
Dahyun Chung, Siyoon Jin, Hyunwook Choi, Honggyu An, Junyoung Seo, Hyunsung Kim, Seung Wook Kim, Seungryong Kim
KAIST
Published October 8, 2026 · on the feed October 9, 2026 · ▲ 57 · arXiv 2610.12299
The abstract
Egocentric world models predict first-person observations conditioned on an agent's actions, but most focus on a single agent. Real embodied settings often involve multiple agents that act and interact within a shared environment. Existing multi-agent world models rely on coarse actions like locomotion, camera control, or discrete commands, leaving fine-grained embodied interactions underexplored. We formulate multi-agent egocentric world modeling as synchronized ego-stream generation for multiple agents interacting through fine-grained actions in a shared world. This requires cross-view action consistency, shared-environment consistency, and consistent propagation of interaction-induced state updates. We propose Multi-agent Egocentric World Model (ME-World), which jointly denoises multiple ego streams in a shared token sequence, conditions each stream on all agents' target-view poses, and grounds generation with shared environment memory. We train and evaluate on real and synthetic multi-agent data and introduce shared-world consistency metrics for environment, update, and identity consistency. Experiments show ME-World improves shared-world consistency, action control, identity preservation, and video quality over existing methods.
The Brief
1:52A short podcast summary of the paper. Written and voiced by AI.
Transcript
Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction. By Dahyun Chung and seven others, from KAIST. Published on the eighth of October.
An egocentric world model predicts what an agent will see from its own point of view, given the actions it takes. Most such models imagine a single agent. Real embodied settings rarely look like that: several agents act and interact in one shared environment. The multi-agent world models that do exist work with coarse actions such as locomotion, camera control or discrete commands, which leaves fine-grained embodied interaction largely unexplored.
The authors frame the task as synchronized ego-stream generation: producing one first-person video stream per agent, all at once, for agents interacting through fine-grained actions in a shared world. Framed that way, three requirements fall out. Actions must look consistent across views. The shared environment must stay consistent. And when an interaction changes the state of the world, that update has to propagate consistently to everyone who can see it.
Their model, the Multi-agent Egocentric World Model, or ME-World, meets those requirements in three moves. It jointly denoises the ego streams of all agents within one shared token sequence. It conditions each stream on the target-view poses of every agent. And it grounds generation in a shared environment memory.
Training and evaluation use both real and synthetic multi-agent data. The authors also introduce shared-world consistency metrics covering environment, update and identity consistency. In experiments, ME-World improves on existing methods in shared-world consistency, action control, identity preservation and video quality.
The shift is from many separate first-person predictions to one joint prediction of a world seen through many eyes. Once the streams are generated together and share a memory, consistency between agents is something the model is built to keep.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
One agent is easy. Two agents sharing a room is where views start to disagree.
You build a video model that predicts what an agent sees from its own eyes, given its actions. It works for one agent alone.
The road ahead
- 1 · The problem ~7 min A video model that imagines one person's first-person view works alone, but separate views of several interacting people end up disagreeing.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Generate all agents' views together in one sequence, give each view every agent's pose, and ground them in a shared memory of the scene.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Set the goal of one stream per agent, denoise them together, condition each on all poses, then ground them in shared scene memory.
- 6 · The experiment ~6 min You’ll guess: Jointly generating several agents' first-person videos, did ME-World also beat existing methods on plain video quality?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.