BREAKING
Robotics Capital Splits: World Models vs VLAs
Two Paths to General Robots
VLA
VLM-based
●
Immediate policy learning
●
Weak long-horizon planning
●
Pi-0, OpenVLA, GR00T
World Model
Video prediction
●
Prediction, data efficiency
●
High compute cost
●
Cosmos, DreamZero, JEPA
0
x
H100 GPUs for Cosmos run
0
mo
pretraining time
0
B
DreamZero parameters
Demos Dazzle, Production Lags
Neither Alone Is Enough
1
Automated data labeling
↓
2
Embodiment retargeting
↓
3
Physics-grounded models
↓
4
Deployment feedback
Breakthrough Path Still Unclear
AI NEWS BLITZ
Robotics investors are split over which approach delivers the field's ChatGPT moment.