Repository navigation
Risk #3 — Measure IK+plan_path false-positive rate on a hand-labelled set #13
Description
DESIGN.html §10 Risk #3 (medium). A task can pass IK reachability + collision-free pathing and still be physically unsolvable (gripper geometry vs object shape, friction). Need to know how often our solvability proxy is wrong.
Probe
- Hand-author 20 tasks within Franka tabletop spanning easy (centred cube), borderline (off-axis, tilted target), and intentionally-hard (object too wide for gripper, target inside a wall)
- Label each ground-truth solvable/unsolvable
- Run our verifier and compute confusion matrix
- If false-positive rate > 20 %: add post-grasp contact-stability check (cube must remain in contact during lift); re-measure
Decide
- < 20 % FP: keep oracle as-is
- 20–50 % FP: add contact-stability check
-
50 % FP: reconsider solver design before scaling to M3
Activity