---

name: go_explore_return_then_explore_loop description: Execute a bounded Go-Explore return-then-explore Phase 1 loop in resettable sparse-reward environments.

Go-Explore Return Then Explore Loop

Use this skill to run the central Go-Explore Phase 1 mechanism after cell encoding and archive update rules are available. It is appropriate for resettable or replayable environments where a selected archive state can be restored before exploratory actions are taken.

Inputs

• A resettable environment exposing state snapshots, actions, and rewards.
• Cell encoder and archive update/select functions.
• Iteration count, rollout horizon, action schedule, and random seed.

Outputs

• Expanded archive.
• Best trajectory, score, and goal flag.
• Trace proving selected cells were restored before exploration.

Workflow

1. Insert the start state into the archive.
2. Select a cell from the archive with seeded deterministic selection.
3. Restore the saved state for that cell without exploratory actions.
4. Execute exploratory actions for a bounded horizon.
5. Encode each visited state and update the archive using score and length rules.
6. Repeat until the iteration budget ends or the goal is reached.

Validation

Run python tests/test_return_loop.py or validate with validate_skill_tree.py --run-tests.

Limitations

This skill is a bounded recovery implementation, not an Atari emulator. It is faithful to the reset/archive/explore mechanism but does not perform neural robustification.