WorldDreamer-V5-Lite 路 RoboCasa
Project page 路 Inference code and installation
WorldDreamer-V5 is a unified world-action model for one-to-many multi-robot collaborative control. Its architecture combines a vision-language backbone, a video expert, a spatial triplane representation, and an action decoder.
This package provides the Lite weights for RoboCasa inference. It uses three camera views and four 14-dimensional state observations to produce 16 steps of 18-dimensional action representations, together with a separate control-mode signal. The inference runtime converts these into the simulator's 12-dimensional commands.
Download
uvx hf download WorldAgents-c/worlddreamer-v5-robocasa \
--local-dir weights/worlddreamer-v5-robocasa
The package contains model.safetensors, checkpoint.json, model_config.json, normalize.json, and the required VLM and video-encoder assets in assets/.
Run
Follow the inference repository README to install the Python 3.10 environments with uv, prepare the text encoder and text cache, and run RoboCasa evaluation. Use the downloaded directory as the checkpoint path. The evaluation split defaults to pretrain.
- Downloads last month
- 12