Add typed configurations to Arena policies

[cvolkcvolk]

Summary

Decouple policy construction from argparse

Detailed description

• Add a generic PolicyCfg contract and limit PolicyBase to typed runtime configs.
• Rename first-party policy configurations consistently with the Cfg convention.
• Preserve the existing CLI through the shared dataclass-to-argparse compatibility adapter.
• Continue the broader migration from argparse namespaces to strongly typed configurations.

[greptile-apps]

Greptile Summary

This PR decouples Arena policy construction from argparse by introducing a PolicyCfg base dataclass and making PolicyBase generic over it, then migrating all first-party policies to the new pattern while keeping deprecated add_args_to_parser/from_args adapters in place as a compatibility bridge.

• PolicyBase is now PolicyBase[PolicyCfgT]; add_args_to_parser, from_args, and from_dict are removed from the base contract. A new register_policy(cfg_type=…) decorator form registers the config type so eval_runner can deserialize Job.policy_config_dict without going through argparse.
• All first-party policy configs (ZeroActionPolicyCfg, RslRlActionPolicyCfg, ReplayActionPolicyCfg, Gr00tRemoteClosedloopPolicyCfg, Pi0RemotePolicyCfg) are renamed to the Cfg suffix and inherit PolicyCfg; the GR00T scheduler and openPI embodiment adapter are promoted into serializable config fields.
• New tests in test_policy_cfgs.py and the updated GR00T/OpenPI test suites verify typed registration, policy construction without deserialization, and scheduler/adapter round-trips through the config.

Confidence Score: 4/5

The change is a well-scoped refactor with a clear compatibility bridge; no existing policy construction paths are broken.

Two spots use assert for runtime validation of user-supplied config data and API-misuse guards — one in the scheduler branch of Gr00tRemoteClosedloopPolicy.init and one in register_policy. Both are silently stripped by Python's optimizer, so an invalid scheduler string from a JSON job config or an incorrect decorator call could produce a confusing downstream failure rather than a clear error message. Everything else in the refactor looks correct and well-tested.

isaaclab_arena/assets/register.py (assert-based API guards) and isaaclab_arena_gr00t/policy/gr00t_remote_closedloop_policy.py (assert-based scheduler validation).

Important Files Changed

Filename	Overview
isaaclab_arena/policy/policy_base.py	Removed argparse abstracts (add_args_to_parser, from_args, from_dict), added PolicyCfg base dataclass and Generic[PolicyCfgT] type parameter — clean simplification.
isaaclab_arena/assets/register.py	register_policy refactored to decorator factory; uses assert statements instead of exceptions for API misuse, which can be silenced by -O.
isaaclab_arena/assets/registries.py	Added _cfg_types dict and register_policy/get_policy_cfg_type to PolicyRegistry; safe because SingletonMeta prevents init from running more than once.
isaaclab_arena/evaluation/eval_runner.py	get_policy_from_job updated to use the registry-based typed-config path with a deprecation warning fallback to the old argparse path.
isaaclab_arena_gr00t/policy/gr00t_remote_closedloop_policy.py	Scheduler selection moved from constructor parameter to config field; uses assert for validating the scheduler value at runtime.
isaaclab_arena_openpi/policy/pi0_remote_config.py	Renamed Pi0RemotePolicyArgs to Pi0RemotePolicyCfg (inherits PolicyCfg); added openpi_embodiment_adapter field to make adapter selection serializable.
isaaclab_arena_openpi/policy/pi0_remote_policy.py	Pi0RemotePolicy now resolves embodiment adapter from config field; removed custom from_dict override that was required by the old adapter-as-constructor-arg pattern.
isaaclab_arena/tests/test_policy_cfgs.py	New test file covering typed config registration, policy construction contract, and deprecation warning for bare @register_policy.
isaaclab_arena_gr00t/policy/config/gr00t_closedloop_policy_config.py	Renamed Gr00tClosedloopPolicyConfig to Gr00tClosedloopPolicyCfg; intentionally does not inherit PolicyCfg since it is an internal YAML-loaded translation config, not a PolicyBase runtime config.

Sequence Diagram

%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
    participant JSON as Job JSON Config
    participant ER as eval_runner
    participant PR as PolicyRegistry
    participant Cfg as PolicyCfg (typed)
    participant Policy as PolicyBase[T]

    JSON->>ER: policy_config_dict
    ER->>PR: get_policy_cfg_type(policy_cls)
    alt Typed path (preferred)
        PR-->>ER: PolicyCfgType
        ER->>Cfg: "PolicyCfgType(**policy_config_dict)"
        Cfg-->>ER: cfg instance
        ER->>Policy: policy_cls(cfg)
    else Deprecated argparse fallback
        PR-->>ER: None
        ER->>ER: DeprecationWarning
        ER->>Policy: policy_cls.from_args(parsed_args)
    end
    Policy-->>ER: policy instance

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
    participant JSON as Job JSON Config
    participant ER as eval_runner
    participant PR as PolicyRegistry
    participant Cfg as PolicyCfg (typed)
    participant Policy as PolicyBase[T]

    JSON->>ER: policy_config_dict
    ER->>PR: get_policy_cfg_type(policy_cls)
    alt Typed path (preferred)
        PR-->>ER: PolicyCfgType
        ER->>Cfg: "PolicyCfgType(**policy_config_dict)"
        Cfg-->>ER: cfg instance
        ER->>Policy: policy_cls(cfg)
    else Deprecated argparse fallback
        PR-->>ER: None
        ER->>ER: DeprecationWarning
        ER->>Policy: policy_cls.from_args(parsed_args)
    end
    Policy-->>ER: policy instance

Loading

_{Reviews (1): Last reviewed commit: "Strengthen typed policy configuration te..." | Re-trigger Greptile}

[arena-review-bot]

🤖 Isaac Lab-Arena Review Bot

Summary

This PR decouples policy construction from argparse: PolicyBase becomes Generic[PolicyCfgT] over a typed PolicyCfg marker, the config_class/from_dict/add_args_to_parser/from_args contract is dropped from the base class, and the policy→config-type association moves into PolicyRegistry via @register_policy(cfg_type=...). Scheduler (GR00T) and embodiment-adapter (OpenPI) selection now live in serializable config fields so CLI and JSON construction stay equivalent. It reads as a clean, deliberate step in the argparse→typed-config migration: deprecation warnings and dated TODOs mark every transitional path, the rename to *Cfg is complete across all consumers, and boundaries hold (extensions depend on core PolicyCfg, not the reverse).

Findings

🔵 Improvement — evaluation/eval_runner.py:111 — The two consecutive if policy_cfg_type is not None: guards in get_policy_from_job are identical; the second can fold into the first for legibility (suggestion inline).

Test Coverage

Good. The new test_policy_cfgs.py verifies typed registration, that the deserialization/argparse surface is gone from PolicyBase, and that bare @register_policy still warns. The GR00T and OpenPI tests were updated to the typed constructor and add registration + scheduler/adapter-selection checks — which exercise the same policy_cls(cfg(...)) path eval_runner now uses. The typed get_policy_from_job path itself and its deprecated fallback warning are only covered indirectly; acceptable given the fallback is transitional.

Verdict

Ship it — one optional readability nit, nothing blocking.

[alexmillane]

LGTM!

One question I have is, in a few places you state that there are compatibility branches to policies that don't use config classes. To me it looks as if everything has been switched over. What still remains?