Grasping Under Uncertainty

Grasp Strategy

Pick a mode and run a trial.

Sensor noise σ 0.90

This Mode

Trials

Success rate

–

Avg time / grasp

–

Belief σ now

–

Compare All Modes

Mode	Succ%	Avg s
Naive	–	–
Robust	–	–
Info-Gather	–	–

How it works

The true object pose is drawn at random each trial and hidden from the controller. A simulated noisy sensor only reports a mean estimate plus a Gaussian belief cloud of radius σ around it — the dashed ghost cube marks that estimate, the scattered points sample the belief.

Naive commits immediately: it grasps tightly at the mean estimate. Whenever the true pose lies outside a narrow tolerance, the fingers close on empty space.

Robust also commits immediately, but widens its gripper aperture to roughly cover the belief cloud (≈2σ), trading grasp precision for a much higher chance the object is still inside the jaws.

Info-Gather spends extra time first: it lowers to the estimate and performs a gentle touch probe. Contact sensing is far less noisy than the original vision estimate, so the belief cloud collapses around a new, tighter estimate — the arm then commits a precise grasp against that updated belief.

tolerance_naive  ≈ fixed, tight
tolerance_robust ≈ clamp(2σ, tight, max)
σ_after_probe    ≈ 0.18 · σ   (touch >> vision)

Sensor noise σ — spread of the initial belief cloud; higher σ punishes naive grasping hardest.
Success rate — fraction of trials where the closed gripper actually contained the object.
Avg time / grasp — real seconds per trial including any probing phase, tracked per mode so speed and reliability can be weighed against each other.

Real-world relevance: this is the core trade-off in POMDP-style manipulation planning — commit to the most-likely state, hedge against the whole belief, or pay a small time cost to actively reduce uncertainty before committing.

True object Belief estimate (ghost) Belief cloud

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.