StarVLA fine-tuned checkpoints for the Robotis SG2 (FFW-SG2)

Fine-tuned checkpoints for the paper "Chunk-Based Diffusion VLA with State-Free Inference for Dual-Arm Humanoid Robot Simulation", comparing state-free and state-conditioned inference on the Robotis SG2 AI Worker in NVIDIA IsaacLab.

Links

• Detailed results and videos: https://rao-sanaullah.github.io/starVLA_results/
• Installation and setup guide: https://rao-sanaullah.github.io/starVLA_setup/
• Checkpoints: checkpoints.zip (direct download)

Checkpoints

checkpoints.zip contains the 30k and 200k step checkpoints of the state-free and state-conditioned variants.

Variant	Checkpoint used in the paper
State-free	200k
State-conditioned	30k

The other checkpoint of each variant (state-free 30k, state-conditioned 200k) is included only for the exploratory cross-checkpoint runs shown on the results page.

License

These weights derive from Qwen2.5-VL-3B-Instruct and are subject to its license (Qwen Research License). Please check it before any commercial use.