Implementation:Danijar Dreamerv3 Parallel Env

Knowledge Sources	DreamerV3
Domains	Reinforcement_Learning, Distributed_Systems, Environment
Last Updated	2026-02-15 09:00 GMT

Overview

Concrete tool for running an individual environment process that communicates with the actor inference server via RPC in distributed DreamerV3 training.

Description

The parallel_env() function in embodied/run/parallel.py deserializes the environment factory, creates the environment instance, connects to the actor server via portal.Client, and runs an infinite step loop. On each step, it sends the observation (with envid and is_eval tags) to the actor and receives an action. On disconnection, it reconnects and resets the episode.

Environment 0 additionally creates a logger client to report FPS and usage statistics.

Usage

Spawned as a separate process by combined() for each training and evaluation environment. Can also be launched independently as config.script == 'parallel_env'.

Code Reference

Source Location

• Repository: dreamerv3
• File: embodied/run/parallel.py
• Lines: L416-473

Signature

def parallel_env(make_env, envid, args, is_eval=False):
    """
    Run a single environment process.

    Args:
        make_env: bytes or Callable(int) -> Env
        envid: int - Unique environment index.
        args: elements.Config with actor_addr, logger_addr, log_every.
        is_eval: bool - Whether this is an evaluation environment.
    """

Import

from embodied.run.parallel import parallel_env

I/O Contract

Inputs

Name	Type	Required	Description
make_env	bytes or Callable	Yes	Factory for environment (cloudpickle-serialized)
envid	int	Yes	Unique environment index
args	elements.Config	Yes	actor_addr, logger_addr, log_every
is_eval	bool	No	Whether this is an evaluation environment (default False)

Outputs

Name	Type	Description
Observations	RPC calls	Sent to actor server via portal.Client.act()
Episode stats	RPC calls	Env 0 sends FPS and usage to logger server

Usage Examples

# Spawned by combined():
import portal
import cloudpickle

for i in range(args.envs):
    portal.Process(parallel_env, cloudpickle.dumps(make_env), i, args)

# Or standalone:
# python dreamerv3/main.py --script parallel_env --replica 0

Related Pages

Implements Principle

• Principle:Danijar_Dreamerv3_Distributed_Environment_Execution