---

name: ppo_recovery_update description: Run a deterministic PPO style clipped scalar update over batched rollout evidence for recovery validation.

PPO Recovery Update

Use this skill when an Isaac Gym recovery needs proof that batched rollout data fed a real trainable update. The skill implements a tiny deterministic PPO-style clipped surrogate over scalar log-probability shifts. Do not use it as a full RL algorithm or as evidence of the paper's full robotics scores.

Inputs

• Old log probabilities, advantages, scalar policy parameter, clip epsilon, learning rate.
• Optional finite-difference epsilon.

Outputs

• Loss before and after, parameter before and after, and optimizer-state-change evidence.

Workflow

1. Compute clipped surrogate loss over the batch.
2. Estimate the scalar gradient by finite differences.
3. Apply one gradient-descent update.
4. Report whether the parameter and loss changed.

Validation

Run python tests/test_ppo_update.py or validate this skill tree with tests enabled.

Limitations

This is a reduced scalar update for mechanism recovery. It is not a replacement for Isaac Gym's full PPO implementation.