Tenon Forge v0.4

Function

Compiles cell CAD and SKU manifest into a randomized USD scene, trains a pick policy on GPU, gates it on held-out SKUs, exports a TensorRT engine to Jetson AGX Thor.

Performance (test cell, Oct 2026)

Cycle time	2.9–3.4 s at 2.1 kg
Held-out success	91% mean over 400 SKUs
Range by family	84–96%, 2 of 24 families in exception routing
Success definition	Lifted, transported and released inside the target bin without regrasp, 2 cm placement tolerance, three attempts maximum.
Throughput	1,140 picks per hour gross at 3.1 s cycle.
Commissioning	52 min median after gate pass, 6 of 24 cells
Exception rate	2 per shift at the pilot customer, 93.4% on that family

Interfaces

Arms	UR10e, FANUC M-10iD, KUKA KR 10 R1100
Grippers	Robotiq 2F-85, Schunk Co-act EGP, vacuum array
Fieldbus	PROFINET, EtherNet/IP
API	REST

Compute

Edge	Jetson AGX Thor, 60 W, 30 Hz
Runtime	TensorRT 10.7, INT8, per-gripper calibration
Training	Isaac Sim, Isaac Lab, Omniverse, 8×H100 80 GB, nvcr.io/nvidia/pytorch:26.08
Planning	cuRobo replan, confidence threshold 0.6

Gates

≥88% held-out success · ≤3.5 s cycle · ≤45 N peak force. A policy that misses any gate does not deploy.

Documented limit

Film under 40 µm: 72% held-out, below the 88% gate. Mitigation: vacuum array.

Out of scope

Loose cable, parts under 8 mm, chrome above 300 lux, deformables above 3 kg, moving conveyors.

Commercial

Pilot	US$ 4,500, one cell, three families, six weeks
Production	US$ 890 per cell per month, annual
New family	US$ 2,400, waived under 88%

Internal measurements on the Córdoba test cell. Not a guarantee of performance on other manifests.

Gate report, build 0.4.118 (PDF) · Deployment report, cell 07 (PDF)