DexForce Releases UNITAS, the First Robot-Task-Native 3D World Model, Topping Multiple Benchmarks

One important foundation for why large language models can scale is that massive amounts of knowledge from diverse sources are unified into the same input-output space—text. DexForce's paper UNITAS brings this question to embodied intelligence: perception, action, and world change have long used different representations, and the model's capacity is consumed by "translation" rather than "understanding." UNITAS's answer is the three-dimensional physical world itself, placing what the robot sees, how it moves, and how the world changes as a result into the same three-dimensional physical representation space.

We first need to distinguish which parts are confirmed facts. The paper's release, the public project page and open-source code, and the 1.7B parameter count are verifiable release facts. The first-of-its-kind positioning as "the world's first robot-task-native 3D world model," however, is explicitly labeled in the source as based on the paper authors' own survey, making it the authors' own claim and lacking external confirmation.

The core question this article seeks to pursue is: to what extent is UNITAS's "3D-native" claim actually supported by the evidence it itself provides, and to what extent does it remain within the vendor's narrative?

The source classifies past methods as 3D-aware World Action Models: they introduce depth, point clouds, or geometric constraints, but their core representation is still built on 2D vision or other latent features. The difference UNITAS claims is that three dimensions are no longer merely auxiliary information but the foundation for the model's understanding of the world—the backbone model directly processes 3D tokens, without using image tokens as a bridge.

Concretely, this claim maps onto the joint modeling of action flow and scene flow. Action flow refers to the motion trajectories of points on the manipulated body, and scene flow refers to the motion trajectories of points in the scene; the two are connected in the same coordinate system, making robot control and world prediction a problem within the same coordinate system. The paper's ablation experiments claim that after adding scene-flow supervision, the performance of the action policy itself continues to improve.

To support joint learning across embodiments and datasets, UNITAS proposes two modules: world-aligned 3D positional embeddings, used to unify 3D coordinate encoding and to complete missing depth within the view frustum, thereby supporting both RGB-D and RGB data; and a physical-time trajectory encoder, which uses FiLM modulation to align trajectories of different durations and frequencies into the same physical-time feature space. The backbone model, Point Dynamics DiT, has both a policy mode and a simulator mode.

An uncertainty must be flagged here: the source only provides ablation results and does not conduct external comparisons, so among the gains brought by the "3D-native" representation itself, how much comes from the representation and how much from data and training strategy cannot yet be separated.