Local Vs Remote Policy

praxis-eval supports two policy execution modes.

Local Policy

Local mode runs the evaluator and policy adapter in one process:

from praxis_eval import LocalPolicy

policy = LocalPolicy(my_policy)

my_policy may be a callable or an object with an act(...) method. If it defines reset(...), the adapter calls it before rollout waves.

Local mode is the simplest option when the simulator stack and policy stack can coexist in one Python environment.

Remote Policy

Remote mode uses praxis-remote as an optional transport layer:

from praxis_eval import RemotePolicy

policy = RemotePolicy("127.0.0.1:50051", timeout=30.0)

The remote server receives the same observation mappings and returns the same batched numpy action contract. The transport package is separate from praxis-eval: https://github.com/Chaoqi-LIU/praxis-remote.

Remote mode is useful when:

• the policy runtime and simulator runtime need incompatible dependencies;
• the policy should run on a different host or accelerator;
• a dedicated simulator runtime calls back into a policy server, as with SimplerEnv and MS-HAB.

Install the remote extra only when needed:

pip install "praxis-eval[remote]"

praxis-remote is optional transport, not a required part of the core evaluator.