vls_calvin_dp_base

A Diffusion Policy checkpoint trained on the CALVIN benchmark, released as the frozen base policy used in VLS: Steering Pretrained Robot Policies via Vision-Language Models.

VLS is a training-free, inference-time framework that steers the sampling process of a frozen generative robot policy (such as this checkpoint) using trajectory-differentiable rewards synthesized by a vision-language model — no fine-tuning or weight updates required.

• Paper: https://arxiv.org/abs/2602.03973
• Project page: https://vision-language-steering.github.io/webpage/
• Trained with LeRobot.