griddly-2d-embodied-inventory

name

griddly-2d-embodied-inventory

description

Operate and modify the Griddly 2D Embodied Manipulation With Inventory framework. Use when an AI agent must run the Docker/Ray/Android WootzApp runtime, generate 2D Embodied Manipulation With Inventory tasks from OpenAI or human boards, validate and solve levels through browser CDP, run Gemini eval, compute oracle reward/policy reward/learnability JSON, debug noVNC/CDP/Docker issues, or safely edit the Griddly pipeline code and docs.

Griddly 2D Embodied Manipulation With Inventory Framework

Use this skill to work in the repository root for the 2D Embodied Manipulation With Inventory Griddly pipeline. Prefer the checked-in scripts over ad hoc Docker/CDP commands.

Core Flow

OpenAI board or human board
  -> tasks/2D_Embodied_Manipulation_With_Inventory/<mode>_<timestamp>/
  -> browser CDP validation
  -> browser CDP oracle planner
  -> browser CDP oracle replay
  -> Gemini eval through scripts/run_all_tasks.sh
  -> browser CDP policy ratio and learnability
  -> generated_tasks/2D_Embodied_Manipulation_With_Inventory/latest_scores.json
  -> regret_score/<model>/regret_scores.json
  -> oracle_score/oracle_scores.json
  -> optional Tinker training with --regret-task-set regret_score/<model>/regret_scores.json

First Commands

Run preflight:

./scripts/griddly_doctor.sh --require-openai

Build runtime:

docker compose -f docker-compose.ray.yml --profile standalone build

Start/inspect runtime:

./scripts/griddly_runtime.sh start
./scripts/griddly_runtime.sh status

Run base eval:

./scripts/run_all_tasks.sh --task 2D_Embodied_Manipulation_With_Inventory --model gemini-2.5-pro --visible

Train in regret-task-set order:

./scripts/run_all_tasks.sh --tinker --regret-task-set regret_score/gemini-2.5-pro/regret_scores.json --model Qwen/Qwen3-4B-Instruct-2507

User Shortcut Behavior

When the user says "generate N tasks using AI", create **N accepted AI task folders under tasks/2D_Embodied_Manipulation_With_Inventory/, not top-level task folders and not merely N OpenAI candidates. Use OPENAI_MODEL from .env unless the user names a model. Count only accepted tasks from latest_generation.json; rejected candidates do not count. If OpenAI returns malformed boards, rerun generation until N accepted tasks exist or a real blocker is clear. Save the accepted task IDs to:

generated_tasks/2D_Embodied_Manipulation_With_Inventory/runs/<timestamp>/task_lists/ai_tasks.txt

When the user says "run regret on those N tasks" or "run regret", score only task IDs under 2D_Embodied_Manipulation_With_Inventory/. Score AI and human task sets separately when both are present:

.../runs/<timestamp>/ai_scores.json
.../runs/<timestamp>/human_scores.json
regret_score/<model>/regret_scores.json
oracle_score/oracle_scores.json

Use python3 -m w8_rl.griddly.inventory_cli score with Gemini gemini-2.5-pro unless the user names another policy model. The browser CDP oracle, oracle replay, policy reward, regret score, and learnability must come from the existing CDP/eval flow.

When the user says "train based on curated dataset" or similar, start Tinker training from the persistent regret task set with the Qwen model:

./scripts/run_all_tasks.sh \
  --tinker \
  --regret-task-set regret_score/gemini-2.5-pro/regret_scores.json \
  --model Qwen/Qwen3-4B-Instruct-2507

For generation, regret scoring, and training, always create a timestamped run folder under:

generated_tasks/2D_Embodied_Manipulation_With_Inventory/runs/<timestamp>/

Tee logs into <run>/logs/*.log, keep task lists in <run>/task_lists/, and tell the user the actual output folder. Before running any generation, regret scoring, eval, or training command, print the resolved run configuration in the chat and in the log:

Command: <generate-ai | score-ai | score-human | train-regret-set | eval>
Provider: <OpenAI | Gemini via Vertex AI | Tinker>
Model: <resolved model name>
Task source/list: <path or task ids>
Algorithm: <short algorithm label from .env/code>
Output folder: <run folder or training output dir>
Log file: <tee target>

Before visible browser work, also print the resolved URLs from .env, especially:

noVNC: http://localhost:${HOST_VNC_PORT}
CDP: http://localhost:${HOST_CDP_PORT}/json/version
Ray dashboard: http://localhost:${HOST_RAY_DASHBOARD_PORT}

Always use visible mode for this flow. For scripts/run_all_tasks.sh, pass --visible. For inventory_cli score, set VISIBLE=1 WEB_VNC=true in the command environment so the nested eval command opens noVNC.

Operating Rules

1. Treat browser CDP/WootzApp as the source of truth for generated-level validation, oracle planning, oracle reward replay, policy ratio, and learnability.
2. Do not replace browser-side scoring with Python approximations.
3. Do not hardcode 2D Embodied Manipulation With Inventory object characters in new logic; extract from tasks/2D_Embodied_Manipulation_With_Inventory/griddly.yaml via extract_grid_reward_rules_from_yaml.
4. Use scripts/run_all_tasks.sh for Gemini policy eval so scoring uses the same Docker/Ray/Android path as normal framework eval.
5. Use scripts/griddly_runtime.sh only to start/stop/status the standalone runtime. Use run_all_tasks.sh for actual eval/training.
6. Keep all generated and curated task IDs inside 2D_Embodied_Manipulation_With_Inventory/. Do not use top-level generated task folders for this flow.
7. Always select visible mode for eval/scoring/training so noVNC is available.
8. Keep host ports in .env. Defaults are noVNC 6081, CDP 49224, Ray dashboard 8275.
9. If ports conflict, edit HOST_CDP_PORT, HOST_VNC_PORT, HOST_RAY_DASHBOARD_PORT, and related HOST_* values in .env.
10. Keep generated score JSON compact. Successful score results should not store bulky stdout/stderr tails.
11. regret_scores.json is the training-order file. It is sorted by learnability descending, then regret_score ascending, with unscored rows last.
12. Leave Design2Code/OpenEnv/SkyRL/SWE-bench code alone unless the user asks for broader framework work.

Read References As Needed

• references/fresh-runbook.md: from-scratch setup and the exact 10 OpenAI + human task learnability workflow.
• references/generated-levels.md: task generation, human board rules, validation, scoring, and final JSON fields.
• references/runtime.md: Docker daemon, build, browser APK, runtime, snapshots, eval, and training commands.
• references/code-map.md: files to edit for generation/scoring/runtime work.
• references/troubleshooting.md: common failures and how to debug them.

Validation

For focused checks:

python3 -m pytest \
  tests/griddly/test_embodied_inventory_levels.py \
  tests/lint/test_griddly_embodied_inventory_artifacts.py \
  tests/lint/test_lint_coverage.py \
  -q

For Docker-based focused checks:

docker compose -f docker-compose.ray.yml --profile standalone run --rm --no-deps -T \
  --user root \
  griddly-ray-worker \
  python3 -m pytest \
    tests/griddly/test_embodied_inventory_levels.py \
    tests/lint/test_griddly_embodied_inventory_artifacts.py \
    tests/lint/test_lint_coverage.py \
    -q

Always run git diff --check before finishing edits.