Products that close the loop from twin to trained robot.
Kilnara combines a physics-accurate digital twin, nightly policy retraining, Jetson edge inference, and an agentic operations copilot so robotic lines adapt safely when the factory changes.
Start with one module or run the entire self-improving loop.
Each product is useful on its own and designed to compound when connected with the rest of the platform.
KilnSim
Physics-accurate robotic line twins in OpenUSD and Omniverse.
Explore →KilnTrain
Nightly RL/IL retraining, synthetic data generation, and validation gates.
Explore →KilnEdge
Low-latency perception and policy inference at the cell on NVIDIA Jetson.
Explore →KilnOps
A NeMo-powered copilot for change-impact analysis and operational decisions.
Explore →Build a physics-accurate twin your robots can train inside.
KilnSim creates and maintains an OpenUSD digital twin of each robotic cell, calibrated to real fixtures, robot limits, camera geometry, materials, and cycle-time constraints.
- ✓OpenUSD asset graph — versioned cells, tooling, parts, and fixtures in Omniverse.
- ✓Physics calibration — friction, mass, lighting, force limits, and robot envelopes matched to reality.
- ✓Change detection — compare scans, CAD updates, and telemetry against the live twin.
- ✓Scenario library — capture edge cases once and replay them on every future policy.
Retrain policies overnight with synthetic data and a validation gate.
KilnTrain runs reinforcement learning and imitation learning jobs nightly, generating the data your cell cannot safely collect and promoting only policies that beat your thresholds.
- ✓RL and IL pipelines — combine demonstrations, telemetry, and simulated rollouts.
- ✓Synthetic data factory — generate labeled scenes for lighting, SKU, pose, and occlusion changes.
- ✓Validation gate — enforce cycle time, success rate, collisions, and force bounds.
- ✓Audit-ready promotion — every candidate policy has lineage, metrics, and approval history.
| Rollouts | 42,000 | new SKU mix |
| Best candidate | policy v42 | 98.4% success |
| Gate | passed | 0 collisions |
| Promotion | pending | ops approval |
Deploy perception and policy inference where cycle time matters.
KilnEdge packages validated policies for NVIDIA Jetson devices with TensorRT acceleration, cell-level observability, and rollback controls for production robotics teams.
- ✓Jetson edge runtime — optimized for Orin modules near robots and cameras.
- ✓TensorRT inference — low-latency perception and policy execution.
- ✓Robot controller bridges — ROS 2, PLC, and vendor driver integration.
- ✓Safe rollout controls — canary deploys, hold-to-approve, and instant rollback.
12 ms
Median policy inference latency after TensorRT optimization.
Jetson
Packaged for edge deployment with local failover behavior.
ROS 2
Topics, services, and action bridges for production cells.
Rollback
Return to the last validated policy without reprogramming.
Ask the twin what changed and what it will cost.
KilnOps is a NeMo-based agentic copilot for automation teams, translating natural-language questions into twin queries, simulations, runbooks, and change-impact reports.
- ✓Natural-language impact analysis — ask about SKU, fixture, camera, or rate changes.
- ✓Runbook generation — produce deploy steps and rollback criteria from validated policies.
- ✓Approval workflows — route decisions to robotics, quality, and operations owners.
- ✓Line intelligence — explain anomalies using telemetry, policy lineage, and twin state.
Operator: What happens if we move camera C2 by 40 mm?
KilnOps: Retrain required. Expected pick success drops to 91.8% with current policy. Simulated recalibration + 18k synthetic scenes restores 98.1% by next shift. No fixture collision risk detected.
How the products fit together.
Telemetry keeps the twin current, the twin trains policies, validated policies run at the edge, and operations teams manage the loop in plain language.
Capabilities across the product suite.
Use this as a planning map for pilots, line rollouts, and plant-wide standardization.
| Capability | KilnSim | KilnTrain | KilnEdge | KilnOps |
|---|---|---|---|---|
| OpenUSD / Omniverse twin authoring | ✓ | — | — | View |
| Synthetic data generation | Scene source | ✓ | — | Explain |
| RL / IL policy retraining | — | ✓ | — | Approve |
| Validation gates and audit history | Scenarios | ✓ | Enforce | ✓ |
| Jetson / TensorRT deployment | — | Package | ✓ | Monitor |
| Natural-language change impact | Twin context | Metrics | Telemetry | ✓ |
Kilnara gives our robotics team a common operating model: simulate the change, train safely, deploy at the edge, and explain the decision to operations.
Put your robotic line on a self-improving loop.
Bring a current line challenge and we will show where KilnSim, KilnTrain, KilnEdge, and KilnOps fit.