Researchers pre-trained the robot model InternVLA-A1.5 on 394 recordings of people playing 20 games, about 223 hours, predicting keyboard and mouse events and future visuals.
After 12,400 training updates and the same fine-tuning on robot tasks, success across four LIBERO suites rose from 1,718 of 2,000 (85.90%) to 1,811 of 2,000 (90.55%). Gains by suite were 9.2, 6.8, 0.2 and 2.4 points. Of 40 tasks, 24 improved, 10 got worse and six tied.
The authors note the comparison also included extra training, so it does not show what part of the gain came from game data.
Background
InternVLA-A1.5
Robot AI model
A vision-language-action model for robots that the authors pre-trained on recordings of human gameplay.
