gcp-robo-cloud 0.1.0

One command to train your robot on cloud GPUs.

gcp-robo-cloud

One command to train your robot on cloud GPUs.

gcp-robo-cloud makes it dead simple for robotics developers to run training jobs on GCP GPU instances. No Kubernetes, no Terraform, no infrastructure expertise needed.

pip install gcp-robo-cloud
gcp-robo-cloud launch train.py --gpu a100

That's it. Your training script runs on a cloud A100, and results are downloaded when it finishes.

Why?

Training robotics policies (RL, imitation learning, sim-to-real) requires GPU compute that most developers don't have locally. Getting cloud GPUs currently means wrestling with VM provisioning, Docker, networking, and cost management. gcp-robo-cloud handles all of that.

Features

• One-command launch - gcp-robo-cloud launch train.py --gpu a100
• Auto-containerization - Detects your dependencies, builds and pushes a Docker image automatically
• Framework agnostic - Works with PyTorch, Isaac Lab, MuJoCo, PyBullet, ROS 2, or any Python project
• Cost-conscious - Spot instances by default (60-91% cheaper), auto-shutdown after training, cost estimates
• File sync - Uploads your code, downloads results automatically via GCS
• Simple auth - Uses your existing gcloud credentials
• Python API - Use programmatically: gcp_robo_cloud.launch(script="train.py", gpu="a100")

Quick Start

1. Install

pip install gcp-robo-cloud

2. Authenticate (if you haven't already)

gcloud auth application-default login
gcloud config set project YOUR_PROJECT_ID

3. First-time setup

gcp-robo-cloud config --init

This enables the required GCP APIs (Compute Engine, Cloud Storage, Artifact Registry).

4. Launch training

cd my-robotics-project/
gcp-robo-cloud launch train.py --gpu a100

Available GPUs

GPU	VRAM	Spot $/hr	On-demand $/hr	Best for
t4	16 GB	$0.11	$0.35	Prototyping, small models
l4	24 GB	$0.22	$0.72	Inference, medium training
v100	16 GB	$0.74	$2.48	General training
a100	40 GB	$1.10	$3.67	Large-scale training
a100-80gb	80 GB	$1.47	$4.90	Very large models
h100	80 GB	$3.54	$11.81	Maximum performance

CLI Reference

# Launch training
gcp-robo-cloud launch train.py --gpu a100
gcp-robo-cloud launch train.py --gpu t4 --no-spot           # On-demand instance
gcp-robo-cloud launch train.py --gpu a100 --args "--epochs 100 --lr 0.001"
gcp-robo-cloud launch train.py --gpu l4 --max-duration 2h   # Auto-stop after 2 hours
gcp-robo-cloud launch train.py --gpu a100 --async            # Don't wait for completion

# Job management
gcp-robo-cloud status                   # List all jobs
gcp-robo-cloud status <job_id>          # Job details
gcp-robo-cloud stop <job_id>            # Stop and delete VM

# Cost estimation
gcp-robo-cloud estimate --gpu a100 --duration 2h
gcp-robo-cloud estimate --all --duration 4h     # Compare all GPUs

# Configuration
gcp-robo-cloud config --init            # First-time setup
gcp-robo-cloud config --show            # Show current config
gcp-robo-cloud config --set gpu=a100    # Set defaults

Python API

import gcp_robo_cloud

# Blocking - waits for training to complete
result = gcp_robo_cloud.launch(
    script="train.py",
    gpu="a100",
    args="--epochs 100",
    spot=True,
    max_duration="2h",
)
print(result.output_dir)   # Path to downloaded results
print(result.cost_usd)     # Estimated cost

# Non-blocking - returns immediately
job = gcp_robo_cloud.launch(script="train.py", gpu="a100", wait=False)
print(job.id)              # Job ID for tracking

# Check status later
status = gcp_robo_cloud.status(job.id)
print(status.state)

# Stop a running job
gcp_robo_cloud.stop(job.id)

Configuration

Create a gcp-robo-cloud.yaml in your project root for persistent settings:

project: my-gcp-project
region: us-central1
gpu: a100
spot: true
max_duration: 4h

docker:
  base_image: nvidia/cuda:12.4.1-runtime-ubuntu22.04
  python_version: "3.11"
  system_packages: [libgl1-mesa-glx, libegl1]

sync:
  exclude: [data/raw/, "*.pth", __pycache__/]
  output_patterns: ["outputs/**", "checkpoints/*.pth"]

script: train.py
args: "--num-envs 4096 --headless"

How It Works

1. Detects dependencies - Finds requirements.txt or pyproject.toml
2. Builds Docker image - Generates a Dockerfile, builds locally, pushes to Artifact Registry
3. Uploads code - Syncs project files to GCS (training code is not baked into the image for fast iteration)
4. Provisions VM - Creates a spot GPU instance with Container-Optimized OS
5. Runs training - Pulls the image, downloads code, executes your script
6. Streams logs - Shows training output in your terminal
7. Downloads results - Copies artifacts (models, logs) to your local machine
8. Cleans up - VM self-terminates after training to stop billing

Custom Dockerfiles

If you need more control, place a Dockerfile in your project root. gcp-robo-cloud will use it directly instead of auto-generating one.

Or specify a base image in your config:

docker:
  base_image: nvcr.io/nvidia/isaac-lab:4.5.0

File Sync

By default, gcp-robo-cloud uploads your project files (excluding .git, __pycache__, .venv, etc.) and downloads everything from outputs/ after training.

Create a .gcp-robo-cloud-ignore file (gitignore syntax) to customize:

# Don't upload large data files
data/raw/
*.hdf5
datasets/

# Don't upload pre-trained models
*.pth
*.onnx

Requirements

• Python 3.10+
• Docker (for building images locally)
• gcloud CLI (for authentication)
• A GCP project with billing enabled

License

Apache 2.0