Reinforcement Learning Training

OpenSandbox provides isolated environments for training reinforcement learning agents, enabling safe experimentation with simulations, hyperparameter tuning, and distributed training workflows.

Overview

RL training in OpenSandbox offers:

Isolated training runs - Each experiment runs in a clean container
Reproducible environments - Consistent package versions and configurations
Resource control - CPU, memory, and GPU allocation per training job
Full observability - Capture logs, metrics, and checkpoints
Distributed training - Scale across multiple sandboxes
Dependency isolation - No conflicts between different RL frameworks

Quick Start

1. Start OpenSandbox Server

uv pip install opensandbox-server
opensandbox-server init-config ~/.sandbox.toml --example docker
opensandbox-server

2. Run RL Training Example

import asyncio
import os
from datetime import timedelta
from opensandbox import Sandbox
from opensandbox.config import ConnectionConfig

async def train_rl_agent():
    sandbox = await Sandbox.create(
        "opensandbox/code-interpreter:v1.0.1",
        connection_config=ConnectionConfig(domain="localhost:8080"),
        env={"RL_TIMESTEPS": "5000"},
        timeout=timedelta(minutes=10)
    )
    
    async with sandbox:
        # Install RL dependencies
        await sandbox.files.write_file("requirements.txt", """
        gymnasium
        stable-baselines3
        tensorboard
        """)
        
        result = await sandbox.commands.run(
            "python3 -m pip install -r requirements.txt"
        )
        
        # Upload training script
        training_script = load_training_script()  # See example below
        await sandbox.files.write_file("train.py", training_script)
        
        # Run training
        train_result = await sandbox.commands.run("python3 train.py")
        
        # Get results
        summary = await sandbox.files.read_file("training_summary.json")
        print(summary)
        
        await sandbox.kill()

asyncio.run(train_rl_agent())

View the complete example: examples/rl-training/

Training Script Example

The example uses Stable-Baselines3 to train a DQN agent on CartPole:

import json
import os
import gymnasium as gym
from stable_baselines3 import DQN
from stable_baselines3.common.evaluation import evaluate_policy

# Configuration from environment
timesteps = int(os.getenv("RL_TIMESTEPS", "5000"))
tensorboard_log = os.getenv("RL_TENSORBOARD_LOG", "runs")

# Create environment
env = gym.make("CartPole-v1")

# Create DQN model
model = DQN(
    "MlpPolicy",
    env,
    verbose=1,
    tensorboard_log=tensorboard_log,
    learning_rate=1e-3,
    buffer_size=10000,
    learning_starts=1000,
    batch_size=32,
    train_freq=4,
    gradient_steps=1,
)

# Train the agent
model.learn(total_timesteps=timesteps)

# Save checkpoint
os.makedirs("checkpoints", exist_ok=True)
checkpoint_path = "checkpoints/cartpole_dqn"
model.save(checkpoint_path)

# Evaluate
mean_reward, std_reward = evaluate_policy(model, env, n_eval_episodes=5)

# Save summary
summary = {
    "timesteps": timesteps,
    "mean_reward": float(mean_reward),
    "std_reward": float(std_reward),
    "checkpoint_path": f"{checkpoint_path}.zip",
}

with open("training_summary.json", "w") as f:
    json.dump(summary, f, indent=2)

print("Training summary:", summary)
env.close()

Use Cases

Hyperparameter Tuning

Run parallel experiments with different hyperparameters:

import asyncio
from typing import List, Dict

async def tune_hyperparameters(configs: List[Dict]):
    tasks = []
    
    for i, config in enumerate(configs):
        task = train_with_config(config, run_id=i)
        tasks.append(task)
    
    results = await asyncio.gather(*tasks)
    
    # Find best configuration
    best = max(results, key=lambda r: r["mean_reward"])
    print(f"Best config: {best}")
    return best

async def train_with_config(config: Dict, run_id: int):
    sandbox = await Sandbox.create(
        "opensandbox/code-interpreter:v1.0.1",
        env={
            "LEARNING_RATE": str(config["lr"]),
            "BATCH_SIZE": str(config["batch_size"]),
            "RUN_ID": str(run_id)
        }
    )
    
    async with sandbox:
        # Install deps
        await install_rl_dependencies(sandbox)
        
        # Upload training script
        await sandbox.files.write_file("train.py", training_script)
        
        # Train
        await sandbox.commands.run("python3 train.py")
        
        # Get results
        summary = await sandbox.files.read_file("training_summary.json")
        result = json.loads(summary)
        result["config"] = config
        
        await sandbox.kill()
        return result

# Run tuning
configs = [
    {"lr": 1e-3, "batch_size": 32},
    {"lr": 1e-4, "batch_size": 64},
    {"lr": 5e-4, "batch_size": 128},
]

asyncio.run(tune_hyperparameters(configs))

Multi-Environment Training

Train on multiple environments simultaneously:

async def train_multiple_environments(environments: List[str]):
    tasks = [
        train_on_environment(env_name)
        for env_name in environments
    ]
    
    results = await asyncio.gather(*tasks)
    return results

async def train_on_environment(env_name: str):
    sandbox = await Sandbox.create(
        "opensandbox/code-interpreter:v1.0.1",
        env={"GYM_ENV": env_name}
    )
    
    async with sandbox:
        await install_rl_dependencies(sandbox)
        await sandbox.files.write_file("train.py", training_script)
        await sandbox.commands.run("python3 train.py")
        
        summary = await sandbox.files.read_file("training_summary.json")
        await sandbox.kill()
        return json.loads(summary)

# Train on multiple environments
envs = ["CartPole-v1", "MountainCar-v0", "Acrobot-v1"]
results = asyncio.run(train_multiple_environments(envs))

Checkpoint Management

Save and restore training checkpoints:

async def save_checkpoint(sandbox: Sandbox, checkpoint_name: str):
    # Download checkpoint from sandbox
    checkpoint_data = await sandbox.files.read_file(
        f"checkpoints/{checkpoint_name}.zip",
        binary=True
    )
    
    # Save to external storage
    with open(f"./local_checkpoints/{checkpoint_name}.zip", "wb") as f:
        f.write(checkpoint_data)

async def resume_training(checkpoint_path: str):
    sandbox = await Sandbox.create("opensandbox/code-interpreter:v1.0.1")
    
    async with sandbox:
        # Upload checkpoint
        with open(checkpoint_path, "rb") as f:
            checkpoint_data = f.read()
        
        await sandbox.files.write_file(
            "checkpoint.zip",
            checkpoint_data,
            binary=True
        )
        
        # Resume training
        resume_script = """
import gymnasium as gym
from stable_baselines3 import DQN

env = gym.make("CartPole-v1")
model = DQN.load("checkpoint.zip", env=env)
model.learn(total_timesteps=10000)  # Continue training
model.save("checkpoints/resumed")
        """
        
        await sandbox.files.write_file("resume.py", resume_script)
        await sandbox.commands.run("python3 resume.py")
        
        await sandbox.kill()

TensorBoard Monitoring

Visualize training metrics with TensorBoard:

async def train_with_tensorboard():
    sandbox = await Sandbox.create("opensandbox/code-interpreter:v1.0.1")
    
    async with sandbox:
        await install_rl_dependencies(sandbox)
        
        # Start TensorBoard in background
        await sandbox.commands.run(
            "tensorboard --logdir runs --host 0.0.0.0 --port 6006",
            opts=RunCommandOpts(background=True)
        )
        
        # Get TensorBoard endpoint
        tb_endpoint = await sandbox.get_endpoint(6006)
        print(f"TensorBoard: http://{tb_endpoint.endpoint}")
        
        # Run training
        await sandbox.files.write_file("train.py", training_script)
        await sandbox.commands.run("python3 train.py")
        
        # Keep sandbox alive to view TensorBoard
        await asyncio.sleep(600)  # 10 minutes
        await sandbox.kill()

Custom RL Frameworks

Train with different RL libraries:

Ray RLlib

rllib_requirements = """
ray[rllib]
tensorflow
torch
"""

rllib_script = """
import ray
from ray import tune
from ray.rllib.algorithms.ppo import PPOConfig

ray.init()

config = PPOConfig().environment("CartPole-v1")
algo = config.build()

for i in range(10):
    result = algo.train()
    print(f"Iteration {i}: reward={result['episode_reward_mean']}")

algo.save("checkpoints/rllib_ppo")
ray.shutdown()
"""

CleanRL

cleanrl_requirements = """
gymnasium
torch
tensorboard
"""

cleanrl_script = """
# Use CleanRL's single-file implementations
import gymnasium as gym
import torch
import torch.nn as nn
import numpy as np

# DQN implementation from CleanRL
# ... (simplified for brevity)
"""

Environment Configuration

Environment Variables

sandbox = await Sandbox.create(
    "opensandbox/code-interpreter:v1.0.1",
    env={
        "RL_TIMESTEPS": "10000",
        "RL_TENSORBOARD_LOG": "runs",
        "LEARNING_RATE": "1e-3",
        "BATCH_SIZE": "64",
        "GAMMA": "0.99",
        "GYM_ENV": "CartPole-v1"
    }
)

Resource Allocation

sandbox = await Sandbox.create(
    "opensandbox/code-interpreter:v1.0.1",
    memory_limit="4Gi",
    cpu_limit="4",
    timeout=timedelta(hours=2)
)

GPU Support

# Use GPU-enabled image
sandbox = await Sandbox.create(
    "opensandbox/code-interpreter-gpu:v1.0.1",
    gpu_count=1,
    memory_limit="8Gi"
)

Supported RL Frameworks

Stable-Baselines3

pip install stable-baselines3[extra]

Supports: DQN, A2C, PPO, SAC, TD3, DDPG

Ray RLlib

pip install ray[rllib]

Supports: PPO, DQN, APEX, IMPALA, A3C, DDPG, TD3, SAC

TF-Agents

pip install tf-agents

Supports: DQN, DDPG, TD3, SAC, PPO, REINFORCE

CleanRL

pip install cleanrl

Single-file implementations of popular algorithms

Gymnasium Environments

# Classic control
env = gym.make("CartPole-v1")
env = gym.make("MountainCar-v0")
env = gym.make("Acrobot-v1")

# Atari
env = gym.make("ALE/Pong-v5")
env = gym.make("ALE/Breakout-v5")

# MuJoCo
env = gym.make("HalfCheetah-v4")
env = gym.make("Ant-v4")

Performance Optimization

Vectorized Environments

from stable_baselines3.common.vec_env import SubprocVecEnv
from stable_baselines3 import PPO

def make_env():
    def _init():
        env = gym.make("CartPole-v1")
        return env
    return _init

num_envs = 4
env = SubprocVecEnv([make_env() for _ in range(num_envs)])
model = PPO("MlpPolicy", env, verbose=1)
model.learn(total_timesteps=10000)

Parallel Training

async def parallel_training(num_runs: int):
    tasks = [
        train_agent(seed=i)
        for i in range(num_runs)
    ]
    
    results = await asyncio.gather(*tasks)
    return results

async def train_agent(seed: int):
    sandbox = await Sandbox.create(
        "opensandbox/code-interpreter:v1.0.1",
        env={"SEED": str(seed)}
    )
    
    async with sandbox:
        # Training with specific seed
        await sandbox.files.write_file("train.py", training_script)
        await sandbox.commands.run("python3 train.py")
        
        summary = await sandbox.files.read_file("training_summary.json")
        await sandbox.kill()
        return json.loads(summary)

Checkpointing Strategy

# Save checkpoints periodically
for i in range(10):
    model.learn(total_timesteps=1000)
    
    # Save checkpoint
    model.save(f"checkpoints/step_{i * 1000}")
    
    # Evaluate
    mean_reward, _ = evaluate_policy(model, env, n_eval_episodes=5)
    print(f"Step {i * 1000}: {mean_reward}")

Troubleshooting

Dependency Installation Failed

Use the Python environment helper:

def with_python_env(command: str) -> str:
    return (
        "bash -lc '"
        "source /opt/opensandbox/code-interpreter-env.sh "
        "python ${PYTHON_VERSION:-3.14} >/dev/null "
        "&& "
        f"{command}"
        "'"
    )

# Install with proper environment
await sandbox.commands.run(
    with_python_env("python3 -m pip install stable-baselines3")
)

Out of Memory

Increase memory limit:

sandbox = await Sandbox.create(
    "opensandbox/code-interpreter:v1.0.1",
    memory_limit="8Gi"  # Increase from default
)

Training Timeout

Increase timeout or reduce timesteps:

sandbox = await Sandbox.create(
    "opensandbox/code-interpreter:v1.0.1",
    timeout=timedelta(hours=4),  # Longer timeout
    env={"RL_TIMESTEPS": "50000"}  # Or fewer timesteps
)

GPU Not Available

Verify GPU support:

result = await sandbox.commands.run("python3 -c 'import torch; print(torch.cuda.is_available())'")
for line in result.logs.stdout:
    print(line.text)

Best Practices

1. Use Ephemeral Sandboxes

# Create fresh sandbox for each run
async def run_experiment():
    sandbox = await Sandbox.create("opensandbox/code-interpreter:v1.0.1")
    try:
        # Training code
        pass
    finally:
        await sandbox.kill()  # Always cleanup

2. Log Everything

# Capture all outputs
result = await sandbox.commands.run("python3 train.py")

for line in result.logs.stdout:
    print(f"[stdout] {line.text}")

for line in result.logs.stderr:
    print(f"[stderr] {line.text}")

if result.error:
    print(f"[error] {result.error.name}: {result.error.value}")

3. Save Artifacts

# Save checkpoints, logs, and metrics
checkpoint = await sandbox.files.read_file("checkpoints/model.zip", binary=True)
summary = await sandbox.files.read_file("training_summary.json")

# Save to external storage
with open("./artifacts/model.zip", "wb") as f:
    f.write(checkpoint)

with open("./artifacts/summary.json", "w") as f:
    f.write(summary)

4. Set Reproducible Seeds

training_script = """
import random
import numpy as np
import torch

seed = int(os.getenv("SEED", "42"))
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)

# Training code...
"""

RL Training Example

Complete RL training example with DQN

AI Coding Agents

AI agents for code generation

Python SDK

SDK reference documentation

API Reference

Full API documentation

​Overview

​Quick Start

​1. Start OpenSandbox Server

​2. Run RL Training Example

​Training Script Example

​Use Cases

​Hyperparameter Tuning

​Multi-Environment Training

​Checkpoint Management

​TensorBoard Monitoring

​Custom RL Frameworks

​Ray RLlib

​CleanRL

​Environment Configuration

​Environment Variables

​Resource Allocation

​GPU Support

​Supported RL Frameworks

​Stable-Baselines3

​Ray RLlib

​TF-Agents

​CleanRL

​Gymnasium Environments

​Performance Optimization

​Vectorized Environments

​Parallel Training

​Checkpointing Strategy

​Troubleshooting

​Dependency Installation Failed

​Out of Memory

​Training Timeout

​GPU Not Available

​Best Practices

​1. Use Ephemeral Sandboxes

​2. Log Everything

​3. Save Artifacts

​4. Set Reproducible Seeds

​Related Resources

RL Training Example

AI Coding Agents

Python SDK

API Reference

Overview

Quick Start

1. Start OpenSandbox Server

2. Run RL Training Example

Training Script Example

Use Cases

Hyperparameter Tuning

Multi-Environment Training

Checkpoint Management

TensorBoard Monitoring

Custom RL Frameworks

Ray RLlib

CleanRL

Environment Configuration

Environment Variables

Resource Allocation

GPU Support

Supported RL Frameworks

Stable-Baselines3

Ray RLlib

TF-Agents

CleanRL

Gymnasium Environments

Performance Optimization

Vectorized Environments

Parallel Training

Checkpointing Strategy

Troubleshooting

Dependency Installation Failed

Out of Memory

Training Timeout

GPU Not Available

Best Practices

1. Use Ephemeral Sandboxes

2. Log Everything

3. Save Artifacts

4. Set Reproducible Seeds

Related Resources