Policies trained in simulation, shipped to the cell.
Tenon Forge compiles your cell and your SKU manifest into a USD scene, trains the pick policy on GPU, and deploys it as a TensorRT engine on Jetson Thor. Policy v0.4, first invoice June 2026.
Pilot US$ 4,500. One cell, three SKU families, six weeks, gate report included.
ISO 13849-1 — PL d documentation suppliedCE 2006/42/EC — by integrator
UR10eFANUC M-10iDRobotiq 2F-85Jetson AGX Thor 60 W
The break point is SKU 400, not the robot.
The demo manifest
Vendors validate on twenty SKUs picked to work. Real manifests run 2,000 to 8,000, and the tail is where grippers fail: soft polybags, crushed cartons, nested trays, film under 40 µm.
Every exception is a person
Each failure becomes a manual exception and each exception becomes someone standing at the cell. Picking projects do not fail on kinematics. They fail on the long tail nobody budgeted for.
Three modules, one pipeline.
1. Scenario compiler v0.4
Imports cell CAD and the SKU manifest, emits a USD scene with physically plausible randomization: friction 0.2–0.9, mass ±18%, lighting 200–1,200 lux, Gaussian depth speckle. 4,096 parallel environments per run.
2. Policy trainer v0.4
PPO with imitation warm-start from twenty minutes of teleoperation. Isaac Lab on an 8×H100 80 GB node: 90 GPU-hours and 11 hours of wall-clock per generation, twenty-two generations a month. Every generation is archived by seed.
3. Edge runtime v0.4
TensorRT engine on Jetson AGX Thor at 30 Hz. Falls back to cuRobo replanning when grasp confidence drops below 0.6, then releases to the exception lane. Runs offline; the plant needs no internet.
Measured, not estimated.
| Cycle time | 2.9–3.4 s at 2.1 kg payload, test cell, October 2026 |
|---|---|
| Grasp success | 91% mean held-out, 400 SKUs, internal benchmark 2026-10 |
| 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. |
| Payload | ≤ 3 kg |
| Grippers | Robotiq 2F-85, Schunk Co-act EGP, vacuum array; others on two-week port |
| Arms | UR10e, FANUC M-10iD, KUKA KR 10 R1100 |
| Edge compute | Jetson AGX Thor, 60 W, 30 Hz control loop |
| Interfaces | PROFINET, EtherNet/IP, REST |
| Training stack | Isaac Sim, Isaac Lab, Omniverse, cuRobo, TensorRT, NGC |
What Forge does not do.
| Transparent film under 40 µm | Depth sensors lose the surface. 72% held-out, below the 88% gate. With a vacuum array the same family reaches 89% and clears. |
|---|
| Loose cable or hose | No reliable grasp point. Routed to manual exception. |
|---|---|
| Parts under 8 mm | Jaw geometry is the limit, not the policy. |
| Glossy chrome above 300 lux | Specular blowout on depth. Dim the cell or add structured light. |
| Deformables above 3 kg | Outside payload and outside the trained mass range. |
| Moving conveyors | Scene must be static during a cycle. Conveyor tracking is roadmap, not shipped. |
Six weeks, fixed scope.
| Week 1 | Cell CAD and manifest intake. Baseline capture of 60 SKUs, manual timings. |
|---|---|
| Week 2 | Scene compilation, first three generations. Per-family success report delivered. |
| Week 3 | Gate run on held-out SKUs. Edge runtime installed, safety validation, handover. |
| Week 4–6 | Supervised production. Drift review every 48 h, manifest updates folded into the next generation. |
Median commissioning after gate pass: 52 minutes, six of 24 cells, August–October 2026.
Published, no quote dance.
| Pilot | One cell, three families, six weeks, gate report | US$ 4,500 |
|---|---|---|
| Production | Per cell per month, annual, offline licence | US$ 890 /mo |
| New SKU family | Training plus gate run, waived under 88% | US$ 2,400 |
Included: scene compilation, retraining, edge image, upgrades. Not included: arm, gripper, safety hardware, integration labour, lighting.
Lead time: pilot starts within 15 business days of purchase order. Gate pass to handover: 52 minutes.
One engineer on site in weeks 1 and 3; travel and per diem included in the pilot price.
Below 88% on a new family is not invoiced. The family stays in exception routing until it clears.
Pilot result, regional 3PL.
“Dejamos de comprar robots que solo funcionan en la celda de demo del proveedor. El cambio de referencia pasó de tres semanas de tuning a menos de una hora.”
Regional 3PL, 14 cells
| Before | 3 weeks on-site tuning per changeover, 71% success at SKU 400 |
|---|---|
| After | 41 minutes, 93.4% at SKU 400, 2 exceptions per shift |
Reference call available on request with the head of automation at the 3PL.
Gate report, build 0.4.118 (PDF) · Deployment report, cell 07 (PDF)
Traction
24 cells in production with 3 customers. ARR US$ 256k, first invoice June 2026. 6 more cells on the bench.
| Regional 3PL | 14 cells |
|---|---|
| National logistics operator | 6 cells |
| Parcel distributor | 4 cells |
Scene and manifest are copied to the training node for the duration of a generation and are not retained after it. Only policy artifacts and telemetry counters persist.
- GPU-hours a month, 8×H100 in burst regime
- 1,980
- Compute spend a month
- US$ 4,400
- Generations a month, every one archived by seed
- 22
Eleven people in Córdoba.
Founded November 2023. Seed round of US$ 5.2M, closed October 2026, from two LatAm deep-tech funds and three industrial angels.
Seven more people in engineering and operations, see careers. Everything on this page was built by these eleven.
Office: Av. Juan B. Justo 4650, Piso 2, X5000 Córdoba, Argentina. Bench cells: Parque Industrial Córdoba, Ruta 19 km 4.
Asked on every call.
Will it run on our arm?
UR10e, FANUC M-10iD and KUKA KR 10 are supported today. Other arms need a two-week kinematic port, quoted separately.
How long for a new SKU family?
Five working days including training and gate run. Under 88% the family is not invoiced and stays in exception routing.
Who certifies the cell?
Forge is a software component, not a machine. Integration must satisfy ISO 10218-1 and the cell risk assessment. We supply the PL d documentation and the stop-circuit interface.
Can we see it before paying?
Send six SKUs. We run a free 48-hour bench test on the Córdoba cell and return the per-family success table.
Send six SKUs. Get the success table in 48 hours.
Bench test on the Córdoba cell, free, no commitment. If your tail is above 88%, we will tell you where it breaks.