Real-World Simulator#

Add realistic observation latency and network conditions to your embodied training without leaving simulation.

RLinf provides two independent modules — delay_sampler and net_emulation — each controlled by configuration parameters in your YAML. Both are disabled by default: omit the corresponding parameters and your workflow runs unchanged.

Observation Delay (delay_sampler)#

Emulate per-environment sensor latency by inserting a configurable delay wait via an InsertDelay env wrapper that delays chunk_step and reset returns.

Use this when you need to test how variable observation timing affects policy training, or when you’re preparing a policy for real-world deployment where camera and joint-encoder latencies differ per robot.

Supported distributions:

Type	Parameters	When to use
`constant`	`delay`	Every observation has the same fixed delay.
`uniform`	`min_delay`, `max_delay`	Latency varies uniformly within a known range.
`exponential`	`rate`	Delay follows a Poisson-like arrival pattern.
`gaussian`	`mean`, `stddev`	Realistic sensor jitter with a typical value and spread.

All delay values are in seconds.

env:
  delay_sampler:
    type: uniform
    min_delay: 0.11       # 110 ms
    max_delay: 0.20       # 200 ms

Requirements:

The delay applies only in training mode. Evaluation is unaffected.
The wrapper is applied per-environment-instance via _setup_env_and_wrappers, independent of the runner type (sync / async).

Network Emulation (net_emulation)#

Emulate cross-worker network latency and bandwidth limits with a global scheduler. Setting cluster.net_emulation.enabled launches a NetEmulationManager, one of the scheduler’s global managers, alongside the others on node rank 0. Point-to-point sends and broadcasts between configured worker groups book a transmission slot with it and wait out the resulting delay, so the per-link latency and the per-group bandwidth budget stay consistent cluster-wide. Broadcasts — which is how weight synchronization moves parameters from Actor to Rollout Workers — are charged once on the sender and once per receiving bandwidth group, and the sender waits for the slowest receiver.

Transfers that never leave a node are not emulated: same-device and same-node handoffs go through shared-memory IPC rather than the network.

Use this to test how policies behave under bandwidth-constrained or high-latency links, such as when Env Workers and Rollout Workers are on separate clusters or cloud regions.

Key	Type	Description
`enabled`	`bool`	Toggle network emulation on (`true`) or off (`false`). Default: `false`.
`symmetric`	`bool`	If `true`, every cross-DC pair is mirrored. Default: `true`.
`crossdc_pairs`	`list`	Source-destination pairs with per-pair `delay_ms`.
`bandwidth_groups`	`list`	Endpoint groups sharing a `bandwidth_mbps` budget.

cluster:
  net_emulation:
    enabled: true
    symmetric: true
    crossdc_pairs:
      - src: ["Env:0-1"]
        dst: ["Rollout:0-1", "Actor:0-1"]
        delay_ms: 50
    bandwidth_groups:
      - members: ["Env:0-1"]
        bandwidth_mbps: 1000
      - members: ["Rollout:0-1", "Actor:0-1"]
        bandwidth_mbps: 500

Endpoint names use the GroupName:Rank convention — Env:0 for the first Env Worker, Rollout:1 for the second Rollout Worker. Inclusive rank ranges are supported, so Env:0-3 expands to Env:0 through Env:3. The Group suffix is stripped automatically.

Requirements:

Works with both the synchronous and asynchronous runners.
Independent of delay_sampler. Enable either, both, or neither.

Together#

Combine both modules to simulate a realistic deployment:

env:
  delay_sampler:
    type: uniform
    min_delay: 0.11
    max_delay: 0.20

cluster:
  net_emulation:
    enabled: true
    crossdc_pairs:
      - src: ["Env:0", "Env:1"]
        dst: ["Rollout:0", "Rollout:1"]
        delay_ms: 50

Example#

Full worked example:

examples/embodiment/config/realsimulator_robotwin_adjust_bottle_dagger_openpi.yaml