/Tech18h ago

UMA Model Unifies 3D Motion And Robot Actions For Heterogeneous Learning

6265421411.1M

Original post

Robot Foundation Models need structured yet generic representations. UMA treats 3D object motion as a co-evolving variable with robot actions, bridging heterogeneous data and use cases under a unified sequence modeling paradigm. 👉https://uma-manipulation.github.io

Yunhao Cao@QuantummCookie

Introducing Unified Motion-Action (UMA) Model, a robot foundation model that uses 3D object motion as a shared interface for heterogeneous robot learning. UMA treats motion and action as co-evolving variables, enabling knowledge transfer across data sources and versatile inference. 🧵 1/n

7:44 AM · Jun 19, 2026 · 1.2M Views

Sentiment

Users approve of the UMA model because unifying 3D motion and robot actions by treating them together makes logical sense.

Pos

100.0%

Neg

0.0%

1 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

UMA-MANIPULATION.GITHUB.IOVia

#1420

Posts from X

Most Activity

VIEWS151LIKES1

Aaliya@AIwithAliya

@KuanFang Treating object motion and actions together makes a lot of sense.

16h1511

RETWEETS1

Chao Feng@chaof1234

Our UMA is a unified model for object motion and robot action that learns from heterogeneous data sources using 3D object-motion trajectories as a shared interface. Check it out: https://uma-manipulation.github.io

Yunhao Cao@QuantummCookie

10h3K248

aiueth@aiueth

@KuanFang People underestimate how much of this is just incentive design. Get the incentives right, everything else follows.

15h