DATAFARM reportedly lifts robot-model task success from 8.3% to 56.7%

A DATAFARM researcher says fine-tuning a robot model on raw TAMP-generated trajectories raised task success from 1.7% to 8.3%. Optimizing those trajectories to better match the model’s pretraining data then raised success to 56.7%, near a 60% human-teleoperation oracle baseline, without using human teleoperation. The researcher also reports 85% success on out-of-distribution deformable-manipulation tasks that the team’s current TAMP cannot solve. The code is open-source.