EN

Hugging Face

LeRobot TOPReward

LeRobot TOPReward

Zero-shot reward wrapper using an off-the-shelf VLM token probability.

Specifications

Architecture components: Training-free zero-shot scoring
Backbone: Qwen3-VL-8B-Instruct by default
Frameworks: LeRobot / Transformers
Input modalities: Language instruction
Input modalities: RGB camera images
Languages: English
Model type: Reward model
Pipeline task: Robot reward modeling
Prediction output: Zero-shot task progress score

No published default checkpoint.

Resources

documentation · LeRobot topreward documentation

Submit info