Move the mouse over the stage — that's your teleop command, a desired direction from the tool tip toward the pointer. The arbiter never drives the robot on that raw command alone.
- Intent inference: for each candidate suture point, score how well your motion direction matches the direction to that goal (cosine similarity), then turn scores into a probability with a softmax at the chosen sharpness. That's the confidence bar chart.
- Arbitration: the robot's own vector points at the highest-probability goal. The commanded velocity blends yours and its own by α:
v = (1−α)·v_human + α·v_auto, renormalized.
- Safety envelope: a hard constraint, independent of α — if the blended command would carry the tip into the red no-go structure, the component pointing inward is clipped so the tip slides along the boundary instead. This is the "constrain unsafe actions" idea: autonomy assists, but a low-level safety layer has final veto.
Try α = 0 with high tremor noise (shaky, unassisted manual control), then α ≈ 0.6 (assisted — noise gets smoothed toward the inferred goal), then α = 1 (fully autonomous once intent is confident). Toggle the safety envelope off to see the tip cut straight through the no-go zone instead of sliding around it.