Managing Environments¶

Each call to gym.make("deepracer-v0") creates a Gymnasium environment and, by default, launches a dedicated simulator instance. You can inspect the simulators currently managed by:

import deepracer
deepracer.running()         # list managed simulators on this host

You can also inspect running simulator services directly through your container runtime:

docker ps
podman ps
apptainer instance list

When you are done with the environment, call env.close() to stop and remove the simulator.

Caching¶

If you are creating many environments in one process, simulator startup time can become noticeable. Use cache=True to keep a matching simulator warm after close() so a later environment with the same configuration can reuse it.

env = gym.make("deepracer-v0", cache=True)
env.close()  # keeps the matching simulator warm for reuse in this process

Cleanup¶

A cached simulator is only reused within the same Python process. When you are completely done, or if you want to reclaim resources, stop all simulators managed by the current process:

import deepracer
deepracer.shutdown_all()    # stop simulators owned by this process

If your notebook, script, or job is interrupted, a running simulator can be left behind. In this case, you can use the following cleanup command before starting another batch of environments:

python -m deepracer.clean   # purge *all* running deepracer simulators

This will remove all managed Docker/Podman containers and Apptainer instances.

Resource limits¶

By default, a process can manage up to DEEPRACER_MAX_ENVS=4 simulators at the same time. For each simulator, budget roughly 3 CPUs, 6 GB of memory, and about a minute of startup time.

Note

At present, deepracer is CPU-only, and does not utilize GPU acceleration for simulation.

Parallel rollouts¶

To use multiple environments in "parallel", use SyncVectorEnv. E.g., for three parallel rollouts:

from gymnasium.vector import SyncVectorEnv

def make_environment():
    return gym.make("deepracer-v0")

# spin up 3 environments (should take ~3 mins for boot-up)
envs = SyncVectorEnv([make_environment for _ in range(3)])

Warning

AsyncVectorEnv is not currently supported by deepracer.