Rahil
Co-founder. Ex-ML Researcher, Samsung R&D. Planning and localisation for autonomous vehicles.
zelpi. picks the right model for each step of a task, checks every command before it reaches the robot, and learns from each run.
Four parts, one router.
Pull state-of-the-art open models in one command, or bring your own post-trained model.
zelpi. scores each model on environment, task and server load, and switches before a failure.
Cobots, humanoids, AMRs and quadrupeds join one routed fleet.
Limits are checked before any command reaches the hardware.
Pick a job. zelpi. hands each step to the model most likely to get it right.
The deployment flywheel
Five steps from API key to a robot moving. The code uses π0.5, hosted on zelpi. now.
Join the waitlist and we issue a key for your team, with an allowance per day and per month.
One key covers every model you are cleared for. Treat it like a password: keep it out of shared code and git.
A model doesn't know your robot's limits. Check every target before it moves anything, whichever model sent it.
Clamp every joint target to your robot's range before sending it.
At the start of each session, move toward the first target slowly over a couple of seconds.
If a camera stops updating, stop moving until it recovers.
Whenever the arms are powered, a person can stop them in reach.
Before running a model on an arm it was not built for, measure the tool tip position and set that arm’s own limits. We can help.
Co-founder. Ex-ML Researcher, Samsung R&D. Planning and localisation for autonomous vehicles.
Co-founder. Ex-Supply Chain, Expeditors.
Co-founder. Ex-ML Researcher, Samsung R&D.
Backed by
zelpi. picks the right model for each step of a task, checks every command before it reaches the robot, and learns from each run.
Four parts, one router.
Pull state-of-the-art open models in one command, or bring your own post-trained model.
zelpi. scores each model on environment, task and server load, and switches before a failure.
Cobots, humanoids, AMRs and quadrupeds join one routed fleet.
Limits are checked before any command reaches the hardware.
Pick a job. zelpi. hands each step to the model most likely to get it right.
The deployment flywheel
Five steps from API key to a robot moving. The code uses π0.5, hosted on zelpi. now.
{{stepBody}}
{{stepNote}}
A model doesn't know your robot's limits. Check every target before it moves anything, whichever model sent it.
{{u.p}}
Co-founder. Ex-ML Researcher, Samsung R&D. Planning and localisation for autonomous vehicles.
Co-founder. Ex-Supply Chain, Expeditors.
Co-founder. Ex-ML Researcher, Samsung R&D.
Backed by