Training Runs¶

A training run executes an RL training loop on cloud GPUs using Isaac Sim. You provide a GitHub repo with a gym environment, Canard clones it onto a GPU worker, trains, and uploads results.

Basic Training Run¶

from canard import Client

client = Client(api_url="https://api.canard.cloud")

run = client.submit_training_run(
    name="go2-standing-v1",
    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,
)

Parameters¶

Parameter	Type	Default	Description
`name`	`str`	required	Human-readable run name
`task_name`	`str`	`Template-G1-Training-Direct-v0`	Gym environment ID
`code_url`	`str`	required	GitHub repo with your environment
`code_ref`	`str`	`None`	Branch, tag, or commit (default: repo default branch)
`num_envs`	`int`	`1000`	Parallel environments per GPU
`max_iterations`	`int`	`1500`	Training iterations
`gpu_count`	`int`	`2`	Number of GPUs (1, 2, or 4)
`enable_video`	`bool`	`False`	Record training videos
`video_interval`	`int`	`2000`	Steps between video recordings
`seed`	`int`	`None`	Random seed for reproducibility
`checkpoint_url`	`str`	`None`	S3 URL to resume training from
`wait`	`bool`	`False`	Block until completion

Choosing `num_envs`¶

More environments = more physics work per iteration = faster convergence but longer per-step time.

GPU	Recommended `num_envs`	Reason
RTX 4090 (24GB)	4096	Fills VRAM without OOM
RTX 5090 (32GB)	8192	3.5x throughput vs 4090

Choosing `gpu_count`¶

Multi-GPU provides near-linear speedup for Isaac Sim (physics is embarrassingly parallel).

GPUs	Speedup	Hourly Rate	Best For
1	1.0x	~$0.50/hr	Prototyping, debugging
2	1.7x	~$1.00/hr	Research (default)
4	2.9x	~$2.00/hr	Sweeps, deadlines

PPO Hyperparameters¶

Pass PPO hyperparameters directly:

run = client.submit_training_run(
    name="custom-ppo",
    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,
    learning_rate=3e-4,
    clip_param=0.2,
    entropy_coef=0.01,
    gamma=0.99,
    desired_kl=0.01,
)

See TrainingConfig for all available hyperparameters.

Resume from Checkpoint¶

run = client.submit_training_run(
    name="go2-continued",
    task_name="Template-Go2-Standing-Direct-v0",
    code_url="https://github.com/canard-cloud/go2-standing-env",
    checkpoint_url="s3://isaac-sim-results/runs/run_a1b2c3d4/checkpoints/model_5000.pt",
    max_iterations=10000,
    gpu_count=2,
)

Monitor and Download¶

# Wait with live progress bar
run.wait_for_completion(show_progress=True)

# Check metrics
metrics = run.get_metrics()
print(f"Final reward: {metrics.training.mean_reward:.2f}")

# Download best checkpoint
run.download_best_checkpoint("./checkpoints")

Fault Tolerance¶

If a spot GPU instance is preempted mid-training:

The task is marked as failed on that worker
The task re-enters the queue
Another worker picks it up and restarts

You don't need to handle this — it's automatic.