From instruction to action.
Generate task-conditioned gripper motion and executable robot commands. Direct control can skip flow generation when only actions are needed.
UNITAS / WORLD ACTION MODELING
UNITASA 3D-Native World Action Model for Embodied Manipulation
1 The Chinese University of Hong Kong, Shenzhen2 DexForce Technology Co., Ltd.3 Foshan University4 Sun Yat-sen University5 South China University of Technology
* Equal contributionβ Corresponding author
01 / OVERVIEW
A 3D-native world action model.
02 / EXPLORE IN 3D
A future you can look around.
Drag to orbit. Scroll to zoom.
Play the prediction, or move through time.
Input
Language instruction
Choose an interaction above.
Initial query points Drag to orbit β
Output
Loading scene
β Drag to explore
Loading 3D trajectoriesβ¦
03 / MOTION STUDIES
Many interactions.
A world in motion.
Lifting. Inserting. Opening. Releasing.
Watch predicted scene responses across simulation and real-world interactions.
04 / ARCHITECTURE
A shared representation.
Two complementary capabilities.
From motion to scene evolution.
Condition on a gripper trajectory and predict how queried scene points move over the same physical interval. On this page, the conditioning trajectory is generated by the policy from the same observation and instruction.
Metric groundingWorld-aligned 3D positional embeddings ground visual tokens in a shared physical frame.
Physical timeA trajectory tokenizer encodes each point trajectory as one token, anchored at its current 3D position.
Coupled expertsA point dynamics expert models action and scene flow, while an action expert generates robot commands.
05 / EVALUATION
Strong performance across benchmarks.
99.8%
LIBERO
Average success89.1%
LIBERO-Plus
Overall success93.04%
RoboTwin Randomized
Success rate56.2%
VLABench
Success rate| Method | Params (B) | LIBERO Avg. | LIBERO-Plus | RoboTwin Clean | RoboTwin Rand. | VLABench SR | VLABench PS | VLABench IS |
|---|---|---|---|---|---|---|---|---|
| VLA Policies | ||||||||
| X-VLA | 1.0 | 98.1 | 71.4 | 72.88 | 72.84 | 29.4 | 51.2 | 70.2 |
| VLA-JEPA | 2.0 | 97.2 | 77.9 | β | β | β | β | β |
| Οβ | 3.0 | 94.15 | 53.6 | 65.92 | 58.40 | 29.4 | 44.1 | 55.0 |
| Οβ.β | 3.0 | 96.9 | 85.7 | 82.74 | 76.76 | 48.1 | 62.3 | 64.9 |
| ABot-M0 | 4.0 | 98.6 | 80.5 | 80.42 | 81.16 | β | β | β |
| Video-Based WAMs | ||||||||
| Cosmos Policy | 2.0 | 98.5 | 82.2 | β | β | β | β | β |
| GE-Act | 2.2 | 96.5 | 80.3 | β | β | β | β | β |
| LingBot-VA | 5.3 | 98.5 | 69.5 | 92.93 | 91.55 | β | β | β |
| Fast-WAM | 6.0 | 97.6 | 51.5 | 91.88 | 91.78 | β | β | β |
| Motus | 8.0 | 97.7 | β | 88.66 | 87.02 | β | β | β |
| 3D-Aware Policies | ||||||||
| SpatialVLA | 3.5 | 78.1 | β | β | β | β | β | β |
| LaMP | 5.0 | 98.3 | 79.3 | β | β | β | β | β |
| X-WAM | 6.7 | β | β | 89.8 | 90.7 | β | β | β |
| Dex-BEV | β | 97.8 | β | 76.0 | 42.0 | β | β | β |
| UNITAS | ||||||||
| UNITAS (w/o Pretrain) | 1.7 | 99.5 | 86.5 | 91.08 | 90.74 | 50.3 | 64.3 | 71.2 |
| UNITAS | 1.7 | 99.8 | 89.1 | 92.94 | 93.04 | 56.2 | 67.6 | 69.3 |