LiberAI reports robot learning results using fewer robot demonstrations

LiberAI says its Liber-0 model achieved a progress score of 75.6 using 25% of a robot demonstration pool plus a full glove-recording pool, compared with 72.8 using the full robot pool alone. The eight-task experiment appears in company-sponsored coverage in The Robot Report. The model combines human video, measured hand motion and information about robot bodies to predict future observations and actions. In a separate 60-task evaluation, it achieved 36.0% success across 600 trials without robot demonstrations collected specifically for those tasks. Pretraining still included robot data; success on challenging manipulation and long-horizon tasks was 10.5%.