Runtime model delivery
Swap the model. Not the robot.
Binplex is an edge-to-cloud continuous fine-tuning layer for robotics integrators. Capture edge failure traces, run remote RL fine-tuning loops, and hot-swap LoRA adapters over-the-air—while the line stays running.
Runs on-premise or cloud-connected. Integrates natively with ROS 2 and Edge K3s.
- grasp-detectACTIVE · 12msv4.2.1
- grasp-detectSTAGED · 9msv4.3.0
- weld-seam-trackACTIVE · 18msv2.8.4
- bin-poseIDLEv1.9.0
Deployed across
- Automotive body-in-white
- Food & beverage packaging
- Third-party logistics
- Contract electronics
Platform
Continuous RobotOps for Systems Integrators
You maintain cells you do not own, on factory networks you do not control, facing environmental drift that causes silent field failures. Binplex is built specifically around those three constraints.
- Sub-second OTA swap
In-memory weight hot-swapping
Load fine-tuned LoRA adapters directly into edge GPU VRAM between task execution cycles. Zero controller restarts, zero re-homing, and zero interrupted ROS 2 control loops.
- K3s telemetry daemon
Automated failure trace capture
Lightweight edge sidecar monitors joint torque, frame drops, and grasp completion. Automatically packages multi-modal failure bags when environmental drift is detected.
- Tinker API pipeline
Cloud-native RL fine-tuning
Failure traces stream securely to cloud RL and DPO training pipelines. Fine-tune lightweight vision-action adapters on real physical edge cases in minutes.
- NVIDIA Jetson & x86
ROS 2 & GitOps native
Runs as a containerized K3s sidecar alongside standard ROS 2 nodes. Manage adapter rollouts, versions, and edge deployments using standard GitOps workflows.
- Zero-risk safety
Automatic rollback guardrails
Previous adapter weights remain resident in memory. If task success rates or inference confidence drop below your set threshold, the edge runner reverts instantly.
- Per-cell history
Full physical audit trail
Every trace capture, fine-tuning run, and LoRA promotion is logged with environmental parameters and measured recovery rates—exportable for customer SLA compliance.
How it works
Four steps to continuous robot recovery
- 01
Deploy the edge sidecar
Run a lightweight K3s container pod alongside ROS 2 on your IPC or NVIDIA Jetson. It monitors motor torque, frame rates, and grasp completion in real time.
- 02
Capture failure traces
When physical drift or glare causes a grasp error, Binplex automatically isolates and packages the 10-second multi-modal failure trace into a secure cloud bundle.
- 03
Fine-tune via cloud RL
Your cloud engine runs a specialized DPO/RL fine-tuning loop on the failure trace, generating a fine-tuned LoRA adapter targeted specifically at that edge case.
- 04
Hot-swap over-the-air
Binplex streams the updated LoRA weights to the edge, hot-swapping them into GPU VRAM between execution cycles—restoring task success without stopping the line.
User stories
What integrators do with it
Three patterns we see most often on the floor.
We used to fly an engineer out for every model update. Now a new grasp model reaches nine plants from the office, and the customer sees the audit log before it goes live.
9 plantsupdated from one consoleOperations lead
Robotic cell integrator, 40+ deployed cells
A seasonal SKU change used to mean a weekend of re-teaching. We staged the new bin-pose weights on Thursday, promoted between shifts, and nobody lost a cycle.
0 minunplanned downtimeControls engineer
Food & beverage packaging line
The rollback guardrail paid for itself the first month. A model regressed on a reflective part, throughput dipped, and Binplex reverted before the operator finished reading the alert.
1 cycleto detect and revertAutomation manager
Contract electronics manufacturer
Your next model update should not need a plane ticket.
Pilot Binplex on one line. If it does not remove a trip, you owe us nothing.
Get started
Tell us about your fleet
We onboard a small number of integrators at a time. Send the details below and we will come back with a scoped pilot for one of your lines.