← Index

Vision-Guided Pruning in Isaac Sim

UR5e live RGB-D approach, gated spur release, and home return in a two-tree Blender orchard.

FIG. 02 — 20 s Isaac dashboard: approach, gated release at 7.8 s, measured fall, home return.

Problem

A pruning arm has to approach a thin branch from wrist RGB-D and dual ToF, then release only when tracking and geometry still agree. Jaw occlusion can wipe out image features at the worst moment. A finished recording is not a finished cut, so failed tracking and failed closure stay in the public record.

Approach

The original UR5e and mock pruner run in Isaac Sim 6.0 / Isaac Lab 3 among both source Blender trees, exported to USD with bark maps and collisions. Seeded pyramidal Lucas–Kanade tracking back-projects RTX optical-Z depth into bounded Cartesian commands. Dual 8×8 ray-cast ToF, contact, and geometry gates must pass before a visual-jaw surrogate releases one rigid spur. An independent grader checks the saved capture rather than the renderer’s success label. Branch identity, axis, and radius come from mesh metadata, not a learned detector.

What I built

I built the Isaac capture loop, live RGB-D controller, release gates, independent sequence grader, and the published success and failure recordings. The robot, orchard meshes, and sensor frames come from pinned upstream sources. This repository composes them, measures one known target, and keeps failed runs visible.

Result

The selected success records 200 frames / 20 s, 68 vision commands, and 17/17 sequence checks. One selected spur releases at 7.8 s, falls 809.49 mm, and the arm returns within 0.001 mm of home. A separate failed episode stops at 7.7 s (confidence 0.14165 < 0.15). Discrete rigid-piece release, not wood fracture.

Keep the failure next to the success

Tracking confidence fell below 0.15 at 7.7 s in the failed episode. Closure stopped at 2/3; no release. The public recording is the failed episode, not a cleaned retake.

FIG. 02A — Closure failure at 7.7 s; motion stops without release.

What drives the arm

Live control uses seeded LK features and RTX depth. Farneback flow is an offline diagnostic. Depth is simulator ground truth. The jaw is a visual proxy; there is no blade CAD or wood fracture model.

Stack

Links