[Feature] Report MicroDuck trajectory evaluation metrics

[vmoens]

Summary

Makes complete-trajectory evaluation a MicroDuck environment API instead of implementing environment semantics inside the PPO recipe.

• Adds MicroDuckEnv.trajectory_metrics, which owns the observation layout and diagnostic_* key interpretation.
• Reports tracking, survival, pose, hopping, displacement, world-ground speed, and wrapped heading-rate metrics.
• Excludes padded collector rows from means, rates, paired-step events, and extrema.
• Shrinks examples/microduck/ppo_mujoco.py relative to main; the evaluator now binds the metric callback in one line.

What was wrong and what is fixed

The recipe previously contained more than one hundred lines that decoded MicroDuck-specific observation indices and diagnostic keys. Besides obscuring the PPO example, this made the environment contract look like application boilerplate:

# Before: environment semantics implemented in the training recipe.
def microduck_metrics(trajectories, *, jumping=False):
    velocity = trajectories["next", "observation"][..., 6:8]
    # ... tracking, contacts, hopping, pose, displacement and heading logic ...
    return metrics

Evaluator(
    env,
    actor,
    metrics_fn=partial(microduck_metrics, jumping=jumping),
)

# After: the environment owns its trajectory interpretation.
Evaluator(
    env,
    actor,
    metrics_fn=partial(MicroDuckEnv.trajectory_metrics, jumping=jumping),
)

Padding also cannot be replaced with zero when computing extrema. For example, every observed height may be negative:

mask:                 [ true,  true, false]
height_gain:          [-1.0, -2.0,  99.0]

wrong, pad with 0:    [-1.0, -2.0,   0.0] -> max =  0.0  # never observed
fixed, pad with -inf: [-1.0, -2.0,  -inf] -> max = -1.0  # valid maximum

The same complete-trajectory mask is used for rates, means, and paired-step events.

Tests

• test_microduck_trajectory_metrics_exclude_padding
• test_microduck_example_recurrent_ppo_trains_on_whole_episodes

Both pass locally. The focused regression covers padded garbage, negative valid maxima, wrapped headings, world-ground velocity, and a one-step trajectory.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/rl/4396

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.