GaussianDream++
Efficient 3D Gaussian world modeling for robotic manipulation.
GaussianDream++ is an efficient 3D Gaussian world modeling framework for robotic manipulation.
The project introduces World State Tokens and World Prediction Tokens with a coupled Gaussian world head. The model learns current physical structure and short-horizon environmental change using RGB, depth, alpha, and 3D motion supervision during training. At inference time, Gaussian decoding and rendering modules are removed.
Highlights:
- 3D Gaussian world modeling inside VLA policies.
- Static-dynamic residual transition for local interaction changes.
- 98.6% average success rate on LIBERO.
- Stronger spatial generalization on LIBERO-Plus and real-robot scenes.