Intelligence for the next billion robots

pip install relay-intelligence
Get StartedRead the Docs

Works with your stack

  • PyTorch
  • JAX
  • ROS 2
  • MuJoCo
  • Isaac Sim
  • LeRobot

Platform

Train, evaluate, and serve.

From one GPU to thousands, without the infra

Fine-tune foundation policies or train from scratch on elastic H100 and B200 clusters. Billed per second, scaled to zero when idle.

train/lossstep 0

Capabilities

Everything between the robot and the GPU.

closes the loop between the data your fleet collects and the policies it runs.

01

Elastic GPUs

H100s and B200s on demand. Pay per second and scale to zero when you're done.

02

Edge runtime

Quantize and compile policies for Jetson and x86 targets.

03

Versioned everything

Datasets, checkpoints, and configs tracked end to end.

04

Fleet observability

Per-robot latency, interventions, and success rates in one place.

05

Enterprise ready

SSO, audit logs, VPC peering, and on-prem deployment for teams with strict security needs.

Developers

One client. Any policy.

Send camera frames, robot state, and an instruction; get actions back. Run preprocessing in the cloud or on the robot, for pi0, pi0.5, and GR00T.

pip install relay-intelligence
Read the Docs
infer.pyPython
1from robot.serving.serving_client import ServingClient
2
3client = ServingClient("pi05")
4
5# Camera frames, joint state, and an instruction in; actions out
6actions = client.infer(
7    instruction="pick up the red cup",
8    image={"wrist": wrist_rgb, "base": base_rgb},
9    state=joint_positions,
10)
11
12# Or preprocess on the robot and send only embeddings
13actions = client.infer(
14    instruction, image, state, preprocess="device"
15)

Ship your first policy this afternoon.

Point at your data and deploy. No cluster to manage.

Get StartedView Pricing