1.7B Parameters, Five Benchmarks Conquered: A World Model That Treats 3D as Its Native Language

A room contains a table, drawers, and a pill box. For a robot to find the pills, it first needs to know what dimensionality of world it is looking at.

World action models of the past two years have taken a detour around this problem. They leverage pre-trained video generators to advance frames one by one, indirectly learning physics by predicting the next image. This works, but images are projections—the distance between pixels is not the physical distance. A robot arm may occupy only a few pixels on screen while spanning half a meter in the real world.

On October 8, a 1.7B-parameter model offered a different answer.

Packing Observations, Actions, and World Changes into One Coordinate System

UNITAS puts all three—what the robot sees, how it moves, and how the world changes—into a single 3D coordinate system with metric units. Input and output are no longer two separate things, but one.

Three components support this interface. Action flow represents human hands and robot grippers uniformly as 3D point trajectories. Scene flow describes the displacement of scene points conditioned on those trajectories. A trajectory tokenizer compresses each point trajectory into a single token, anchored at its current 3D position.

The last piece is crucial: because one trajectory occupies one token, the robot can fit more action steps into a limited context window without enumerating every joint at every timestep.

| Component | Role | What it replaces in the traditional pipeline | |---|---|---| | Action flow | Unifies human hands and grippers into 3D point trajectories | Joint angle sequences plus camera poses | | Scene flow | Gives scene point displacement conditioned on actions | Next-frame image prediction | | Position embedding | Anchors visual tokens to 3D coordinates | Projections dependent on depth input | | Trajectory tokenizer | Compresses each trajectory into one token | Per-timestep action sequences |

1.7B Beats 6B and 8B

The most striking comparison in the paper is a reversal of parameter counts.

UNITAS achieves 99.8% average success on LIBERO and 89.1% on LIBERO-Plus. On RoboTwin 2.0's dual-arm manipulation tasks, it reaches 92.94% in the clean setting and 93.04% in the randomized setting. The models it outperforms—Fast-WAM and Motus—have 6B and 8B parameters respectively.

On real robots across six tasks, the average success rate is 85%. On action-conditioned scene prediction, the paper reports up to 49% lower displacement error than PointWorld.

The authors attribute this to three differences: 3D as a native representation rather than an add-on, a cross-embodiment shared interface, and compact trajectory-as-one-token encoding. Only when all three hold can a small model stand up to larger ones.

A Shared Interface Means Hands Don't Need to Relearn

UNITAS is claimed to be the first world action model in which different robots and human hands share the same interface. Technically, this means the action-flow side no longer distinguishes a multi-finger gripper from a five-finger hand—shape differences are moved into local appearance adaptation.

On action-conditioned scene prediction in RoboTwin, the 1.7B UNITAS achieved the best results among compared methods, with displacement error up to 49% lower than PointWorld.

The paper is 27 pages with 11 figures and 13 tables; the code is open-sourced at GitHub under DexForce/UNITAS. Its current status is under review.

Benchmarks Refreshed — At What Cost

| Benchmark | Unit type | UNITAS result | |---|---|---| | LIBERO | Single-arm manipulation | 99.8% avg. success | | LIBERO-Plus | Perturbation robustness | 89.1% | | RoboTwin 2.0 clean | Dual-arm manipulation | 92.94% | | RoboTwin 2.0 randomized | Dual-arm manipulation | 93.04% | | Real robots | Six tasks | 85% avg. |

The real-world 85% trails the simulation results by a noticeable margin—a gap typical of this line of work, and the number most worth watching. Every simulation benchmark shares the same assumption: object mass, friction coefficients, and joint clearances are known and stable. The real world offers no such guarantees.

The "under review" status suggests the authors themselves do not expect these numbers to translate directly into deployment.

Sources and Caveats

• All figures are taken from the arXiv:2610.12099 abstract page and the official project technical report; the experiments section and all 13 tables were not checked item by item.
• The specific LIBERO-Plus and RoboTwin numbers come via a Chinese-language report; the abstract only prints LIBERO 99.8%, real-task 85%, and up to 49% lower displacement error than PointWorld.
• The comparison against Fast-WAM 6B and Motus 8B comes from the project technical report; whether they were measured on the same hardware and protocol was not verified.
• The code is open-sourced; whether the weights were released alongside it was not stated.