One Demonstration.
Every Arm. Confidence-Validated.
Fleet Learning is the architecture underlying everything Automcore does: one operator-guided demonstration on a single arm generates a compressed grasp model that propagates to every other arm controller on your facility LAN — with per-arm confidence validation before any arm uses the model in production.
How It Works
From one arm to the entire fleet
The propagation cycle is designed around your operational schedule — training during production hours, distribution during maintenance windows, validation before production resumes.
PHASE 01 — CAPTURE
3D pose capture on the training arm
An operator guides a single arm through the pick sequence for a new SKU. Automcore's vision system builds a point cloud of the object, estimates stable grasp poses across orientation variance, and records force-torque data from the training arm's end effector. The resulting grasp model encodes pick approach, gripper force envelope, and stacking sequence context — not just waypoints.
PHASE 02 — COMPRESS
Knowledge distillation and arm-family normalization
The captured grasp model goes through a knowledge distillation step that reduces transfer size by 94% while preserving pick performance. The compressed model is then normalized across the target arm families in your fleet — translating joint-space and Cartesian parameters so a model trained on a UR10e can be correctly interpreted by a FANUC CR-14iA or ABB IRB 2400 endpoint. Each normalized package is versioned and signed before distribution.
PHASE 03 — PROPAGATE AND VALIDATE
Scheduled fleet distribution with per-arm confidence gating
During the scheduled maintenance window, the edge node distributes each arm's normalized model package over the facility gigabit LAN. Distribution is staggered by arm to avoid saturating the switch. Each receiving arm runs the model through a local dry-run simulation, compares grasp confidence against the configured threshold, and either activates or flags for remediation. Fleet Manager shows arm-by-arm status in real time throughout the propagation window.
Capabilities
What fleet learning enables
Mixed-Brand Fleet Support
Propagation works across mixed-manufacturer fleets. A training session on a UR10e generates a normalized model that distributes correctly to FANUC CR-14iA and ABB IRB 2400 endpoints in the same fleet. Automcore handles the joint-space and Cartesian translation between arm families — the operator runs one session regardless of how many different arm brands are in the fleet.
Incremental Updates
When a trained SKU improves (better grasp strategy discovered), only the delta update propagates — not the full model. Fleet stays current without repeated full distributions.
Model Versioning
Every model version is tagged and stored. If a new version underperforms, fleet arms can roll back to the previous validated version within minutes.
Offline Resilience
Once a model is distributed, each arm operates independently. The edge node going offline doesn't affect production picking — arms run from their local model cache.
See fleet learning demonstrated against a real facility layout.
We'll walk through how Automcore propagation would work with your specific arm models and facility network.