Add unified policy evaluation for RL checkpoints

[acrlw]

Description

This PR adds embodichain eval-policy as the post-training entry point for EmbodiChain checkpoints and external Policy Profiles.

The change:

• records training configs and best/latest checkpoints in run-manifest.json;
• evaluates a training run Headless with the existing episode evaluator;
• opens the original simulator task with --viewer, using the same observation flattening, deterministic Policy inference, action conversion, reset, and termination behavior as Headless evaluation;
• supports external policies through a registered Motion Profile that provides the model, resources, Adapter, and runtime settings;
• writes each result to evaluations/<timestamp>-policy/evaluation.json; and
• adds a concrete ANYmal-C velocity TorchScript .pt example, Viewer controls, documentation, and focused tests.

flowchart LR
    Run[Training run] --> Manifest[run-manifest.json]
    Manifest --> Runtime[EmbodiChain RL runtime]
    Runtime --> Headless[Headless evaluation]
    Runtime --> Bridge[Task Viewer bridge]
    Profile[External Motion Profile] --> Evaluator[MotionPolicyEvaluator]
    Bridge --> Evaluator
    Headless --> Report[evaluation.json]
    Evaluator --> Report

Loading

Evaluate an EmbodiChain training run:

embodichain eval-policy outputs/<experiment>_<timestamp>
embodichain eval-policy outputs/<experiment>_<timestamp> --viewer --renderer hybrid

Run the external Policy example:

python examples/learning/policy_evaluation/prepare_resources.py
python examples/learning/policy_evaluation/eval_policy.py --viewer --renderer hybrid --sim-device gpu

Viewer controls use Backspace to reset the task and camera, T to switch between tracking and free camera modes, R to start or stop recording, and Esc to close the Viewer. The ANYmal-C example uses W/S for vx, A/D for vy, Q/E for yaw, and M to clear the command.

No linked issue.

Type of change

• Enhancement (non-breaking change which improves an existing functionality)
• New feature (non-breaking change which adds functionality)
• Documentation update

Screenshots

ANYmal-C velocity	CartPole GRPO

Validation

• 79 focused Python tests passed after synchronizing with the latest main.
• All 16 Motion Policy integration tests passed in the isolated evaluation environment.
• PointMass APG Headless evaluation completed and wrote evaluation.json.
• CartPole Hybrid Viewer evaluation completed and wrote evaluation.json.
• The public ANYmal-C .pt completed the selected control steps and wrote evaluation.json.
• Black 26.3.1 and git diff --check passed.

Checklist

• I have run the black . command to format the code base.
• I have made corresponding changes to the documentation.
• I have added tests that prove my fix is effective or that my feature works.
• Dependencies have been updated, if applicable.