StarForce

Dual system training framework for robotics.

The overall structure borrowed from GR00T, with these modifications:

• Simplifer VLM model, introduced more advanced VLM and larger slow system;
• Connecting fast system without cross-attention, using text encoder instead;

Slow thought system

will goes starforce/model/backbone contains various VLMs. Provides a unified interface connect with fast system (action expert)

Fast action system

Currently support:

• DiT: diffusion transformer
• QwenFlow: flowmatching based action expert

Env install

pip install -e .[base]

Training

training scripts goes to scripts/xxx.sh

training fast system:

sh scripts/v0/sl_0.sh

training slow thinking system: