VICTR method raises robot success on unseen tasks from 6.2% to 23.8%

Researchers led by Luca Macesanu have described VICTR in an arXiv paper, a way for vision-language-action robot models to learn from demonstrations in context, without further fine-tuning. Instead of picking demonstration clips by how similar they look, it uses a value function that estimates how far the current task has progressed, judged against a reference video and an instruction.

VICTR then gives a pi-0.5 policy the parts of a demonstration that match the current stage of the task. On real-world manipulation tasks the model had not seen, it lifted success from 6.2% to 23.8% compared with a fine-tuned pi-0.5 model, and task progress from 28.9% to 55.5%, the authors report.