Canard¶
Serverless Isaac Sim RL training. Define your sweep, we handle the GPUs.
Canard is a managed platform for robotics reinforcement learning. Submit training runs and hyperparameter sweeps from Python — Canard distributes them across cloud GPUs, handles fault tolerance, and streams results to a live dashboard.
Install¶
30-Second Example¶
from canard import Client
client = Client(api_url="https://api.canard.cloud")
run = client.submit_training_run(
name="go2-standing",
task_name="Template-Go2-Standing-Direct-v0",
code_url="https://github.com/canard-cloud/go2-standing-env",
num_envs=4096,
max_iterations=5000,
gpu_count=2,
)
run.wait_for_completion(show_progress=True)
run.download_best_checkpoint("./checkpoints")
What You Get¶
| Feature | How |
|---|---|
| Training runs | submit_training_run() — single run on cloud GPUs |
| Hyperparameter sweeps | submit_training_sweep() — grid search across learning rates, entropy, etc. |
| Live monitoring | Dashboard at app.canard.cloud with reward curves |
| Fault tolerance | Spot preemption? Task re-queues automatically |
| Cost estimates | SDK prints estimated cost before you confirm |
| Artifact download | Checkpoints, videos, TensorBoard logs — all in S3 |
| W&B integration | Weights & Biases logging built in |
Training Tiers¶
| Tier | GPUs | Hourly Rate | Best For |
|---|---|---|---|
| Standard | 1x | ~$0.50/hr | Prototyping |
| Fast | 2x | ~$1.00/hr | Research (default) |
| Turbo | 4x | ~$2.00/hr | Sweeps, deadlines |
Next Steps¶
- :material-download: Installation — Install the SDK and set up credentials
- :material-rocket-launch: Quickstart — Submit your first training run
- :material-tune: Parameter Sweeps — Run hyperparameter searches
- :material-book-open-variant: API Reference — Full Client and RunHandle docs