Policies¶
Policies run on the server and are selected together with the algorithm.
Quickstart¶
plugrl-run-server dummy-policy default dummy default
plugrl-run-server dppo-policy default dppo hopper
Verify¶
Confirm the policy UID is visible in the CLI.
Built-in policies¶
dummy-policy: random actions for protocol smoke tests. By default they are continuous, 7-dimensional, with a horizon of 4, as the env client'sdummy-v1expects.fpo-policy: FPO flow-matching policy, used by the get-started rundppo-policy: DPPO diffusion policy. Needs thedppoextra.gaussian-policy: CleanRL's Gaussian MLP, trained withppo.--policy.deterministicacts with the mean instead of a sample, for evaluation witheval.dppo-gaussian-policy: DPPO's Gaussian MLP, which loads the checkpoints DPPO releases. Trained withppo. Needs thedppoextra.pi0-policy: OpenPI policy, requires a checkpoint path. Needs OpenPI.
Which algorithm takes which policy.
fpo: flow policies,fpo-policyandpi0-policy.dppo: diffusion policies and flow policies, since the flow base class subclasses the diffusion one:dppo-policy,fpo-policy,pi0-policy.ppo:gaussian-policy,dppo-gaussian-policy.evalanddummy: any policy.
A policy whose dependencies are missing is also missing from the CLI, with no
error. dppo-policy and dppo-gaussian-policy need the dppo extra
(uv sync --extra dppo in plugrl-server). pi0-policy needs OpenPI, set up
as plugrl-server/src/plugrl_server/policy/openpi/README.md describes.
OpenPI example. pi0-policy runs with eval, with fpo, and with dppo
through its libero variant, which
E25
ran end to end.
plugrl-run-server pi0-policy default eval default \
--policy.name pi05_libero \
--policy.checkpoint-path /path/to/checkpoint \
--policy.device cuda
Troubleshooting¶
- Policy UID not listed: registration module was not imported, or the policy's dependencies are not installed (see above).
pi0-policyfails at startup: complete the OpenPI local setup in the server repo.