An inference router for physical AI. For each step of a robot's task, zelpi. picks the model most likely to succeed, switches before a failure, and logs every run so routing gets better in your environment.
Cobots, humanoids, AMRs and quadrupeds. First deployments are cobot cells on factory lines.
Dream-1, which zelpi. hosts now, plus open vision-language-action models such as π0.5, OpenVLA, GR00T N1 and SmolVLA, and any model you have post-trained.
It looks at the environment, the task and current server load, predicts which model is most likely to succeed at the next step, and switches when another model is the better bet.
Yes. The playground is a MuJoCo physics simulation of a real work cell: two I2RT YAM arms with wrist cameras and scanned everyday objects. The models are real post-trained models (π0.5, ACT, MolmoAct2) running on a GPU, and every call goes through the zelpi. gateway, the same API you would use. Success rates come from closed-loop evaluation in that cell, including task combinations the models never saw in training, and latency is measured live. In a pilot we set it up with your own robots, scenes and models.
You can connect your own model today: choose “Your endpoint” in the playground and its calls go through the zelpi. gateway. To bring your own robot or scene, join the waitlist and tell us your robot and task.
Action guardrails check limits before any command is sent to the hardware. If you run your own control loop, clamp every target to your robot's limits and keep someone on the e-stop while the arms are powered.
Runs are logged and evaluated so routing improves for your environment. Data terms are agreed per deployment.
Factories get a managed cobot cell: we build, install and run it, starting with the station that costs you the most downtime. Robotics teams plug the router into robots they already run.
Pricing is set per pilot. For robotics teams it is usage-based. Join the waitlist and we will reply within two working days.
An inference router for physical AI. For each step of a robot's task, zelpi. picks the model most likely to succeed, switches before a failure, and logs every run so routing gets better in your environment.
Cobots, humanoids, AMRs and quadrupeds. First deployments are cobot cells on factory lines.
Dream-1, which zelpi. hosts now, plus open vision-language-action models such as π0.5, OpenVLA, GR00T N1 and SmolVLA, and any model you have post-trained.
It looks at the environment, the task and current server load, predicts which model is most likely to succeed at the next step, and switches when another model is the better bet.
Yes. The playground is a MuJoCo physics simulation of a real work cell: two I2RT YAM arms with wrist cameras and scanned everyday objects. The models are real post-trained models (π0.5, ACT, MolmoAct2) running on a GPU, and every call goes through the zelpi. gateway, the same API you would use. Success rates come from closed-loop evaluation in that cell, including task combinations the models never saw in training, and latency is measured live. In a pilot we set it up with your own robots, scenes and models.
You can connect your own model today: choose “Your endpoint” in the playground and its calls go through the zelpi. gateway. To bring your own robot or scene, join the waitlist and tell us your robot and task.
Action guardrails check limits before any command is sent to the hardware. If you run your own control loop, clamp every target to your robot's limits and keep someone on the e-stop while the arms are powered.
Runs are logged and evaluated so routing improves for your environment. Data terms are agreed per deployment.
Factories get a managed cobot cell: we build, install and run it, starting with the station that costs you the most downtime. Robotics teams plug the router into robots they already run.
Pricing is set per pilot. For robotics teams it is usage-based. Join the waitlist and we will reply within two working days.