The Plans Worked. The Robot Learned Almost Nothing

Task-and-motion planners can produce successful robot demonstrations at scale, yet DATAFARM’s authors found that feeding raw planner trajectories into a pretrained vision-language-action model transferred surprisingly little. They attribute the gap to behavior: the planner moves through joint configurations, motion styles and timing patterns unlike the data the model saw during pretraining. DATAFARM reshapes the generated demonstrations toward that prior distribution. Across three tabletop tasks and one cloth-folding task, the aligned data produced 56.7% average success, compared with 8.3% for raw planner data and 61.7% for human teleoperation. On an out-of-distribution deformable-object task, the fine-tuned model retained 85% success versus 90% before fine-tuning. These are the authors’ reported tasks and models, not a guarantee that distribution alignment replaces human robot data broadly.