A task-specific policy is a network trained from scratch on demonstrations of one exact task; it has no prior about objects it has never seen, so on a novel object its usable demo count stays near zero against the thousands a from-scratch policy needs. A foundation-model policy is pretrained on demonstrations pooled across many robots and tasks, so a handful of new demonstrations are enough to fine-tune it — few-shot adaptation.
P(success) ≈ cap · min(1, demos_available / demos_required)
task-specific: demos_required ~ 800–3000
foundation: demos_required ~ 4–10
- Approach — switches the policy the arm is running: from-scratch task-specific vs. a pretrained foundation model.
- Try new task — spawns an unfamiliar object shape and resets the demo counters for both.
- Adaptation speed — how fast usable demonstrations accumulate this session.
- Attempt tempo — how often the arm attempts a grasp.
Real-world relevance: this is why generalist robot policies (RT-2, Octo, OpenVLA-style models) can pick up new pick-and-place tasks from a handful of demonstrations, while a narrow single-task controller needs a fresh, large demonstration dataset for every new object or setting.