Pose Estimation to Fleet Propagation.
One Demonstration. Every Arm.
How Automcore captures 3D object pose, constructs a generalized grasp plan, and distributes a compressed model to every arm controller on your facility network — without a cloud round-trip or per-arm teach-in session.
Core Technology
The Propagation Engine
Four technical layers that turn a single operator demonstration into fleet-wide pick capability — from 3D pose capture through per-arm grasp validation.
3D Pose Estimation and Grasp Planning
During the training session, Automcore's vision system builds a point cloud of the SKU's geometry, estimates object pose across relevant orientations, and generates a grasp point distribution — ranking approach angles by estimated pick reliability for the attached end effector (vacuum cup, mechanical gripper, or hybrid). Force-torque data from the training arm calibrates the grasp force envelope for that SKU's weight and surface compliance.
Model Compression via Knowledge Distillation
The full grasp model — pose estimation weights, grasp point rankings, force profiles, and approach trajectory data — is compressed through a knowledge distillation step that retains pick performance while reducing transfer size by up to 94%. The compressed representation includes a version tag and a per-arm validation payload used at the receiving end to confirm model integrity before activation.
Scheduled Fleet Distribution
The Automcore edge node maintains a propagation queue and distributes model packages to each arm controller over the facility's gigabit LAN during the scheduled maintenance window. Distribution is staggered by arm — no simultaneous multi-cast that saturates the switch. Delta-only updates propagate when a model is refined, so arms don't re-download the full package for incremental grasp strategy improvements.
Per-Arm Confidence Gating
Each receiving arm runs the propagated model through a local dry-run simulation before activating it for production. The simulation evaluates grasp confidence against a configurable threshold — typically set at 92–95% for mixed-SKU operations. Arms that pass the threshold activate. Arms below threshold are held and flagged in Fleet Manager, triggering a follow-up training session. No arm is silently put into production below threshold.
Architecture
On-premise.
No cloud round-trips.
Automcore runs entirely within your facility during production picking. The edge node connects to your arms via the facility LAN — no internet required for model propagation, no cloud latency during pick operations.
- Facility LAN with gigabit uplink to arms — no special networking required
- NUC-class edge node (hardware provided) handles all propagation scheduling
- Model data stays on your network — no pick telemetry leaves the facility
- Optional cloud sync for multi-facility deployments (separate configuration)
Ready to see the propagation engine in your facility?
We'll walk through how Automcore would connect to your specific arm models and facility layout.