ProbeNet — Vision-Informed Interactive Perception Link to heading
ProbeNet conditions a manipulation policy on object physical-parameter state (mass, friction, compliance) to generalize across a spectrum of objects with similar appearance but different physics.
Pipeline Link to heading
-
Visual Perception — Vision encoder observes the object → visual params (shape, size, gloss, coarse material class).
-
Vision-Informed Interactive Perception (Probe Phase) — Visual params set the probe’s initial strategy. A scripted probe (squeeze + lift-hold) executes, reading servo current/torque → physical params (mass, compliance, true friction). This step can confirm or correct the visual prior.
-
Manipulation Policy — Conditioned on (visual params + physical params) via a parameter graph / GNN, feeding a BC or ACT action head. Executes pick-and-place with physics-aware trajectory and force.
Headline Experiment Link to heading
Two visually near-identical objects (e.g. matte fragile shell vs. weighted/slippery shell) with opposite correct grip-force/trajectory requirements. Baseline (vision-only) fails asymmetrically — crushes one, drops the other. ProbeNet succeeds on both by resolving the ambiguity through probing.
Hardware Link to heading
- SO-101 leader-follower arm pair
- Static-mounted RealSense D435if (third-person view, RGB stream)
- Servo current/torque readout used as the probe signal source
Object Set Link to heading
- Shell A (fragile): thin-wall PLA print, hollow, low mass
- Shell B (rigid/heavy/slippery): thick-wall or weighted-core print, smoother finish
- Same external geometry (~5-6cm sphere), painted to match