RL Curriculum Orchestrator

RLCurriculumOrchestrator connects the RL domains that were split across the project and the expanded RL domain catalog: Knowledge RL, Gaming RL, game and physics simulation, robot/Gazebo simulation, cybersecurity, Windows operations, software build/test, research, healthcare, tutoring, parenting, finance, stocks, commerce, operations, infrastructure, legal, policy, manufacturing, creative workflows, productivity, IoT, fleet maintenance, call-center support, ROS/robot hooks, and real-world deployment gates.

RL hub (start here): ReinforcementLearning.md

Related subsystem docs:

• docs/RLDomains.md
• docs/MathNN.md — in-process neural policies (rl net, learners)
• docs/LifeRL.md — advisory multi-domain life RL
• docs/KnowledgeRL.md
• docs/GamingRL.md
• docs/DrivingRL.md
• docs/MinecraftPluginTrainingGuide.md
• docs/SimpleAgents.md

The simulator smoke tests are non-actuating. They train tiny tabular policies against safe synthetic RLEnv environments so the RL loop can be checked before anything touches Minecraft, Windows automation, ROS, or hardware.

Neural (MathNN) training is separate: use rl net …, rl cartpole, domain net on flags, or nn math train … after smoke passes.

Commands

rl status
rl curriculum status
rl curriculum plan
rl curriculum next
rl curriculum domains [category]
rl curriculum domain <id>
rl curriculum envs
rl curriculum smoke all [episodes] [max_steps]
rl curriculum smoke knowledge [episodes] [max_steps]
rl curriculum smoke game [episodes] [max_steps]
rl curriculum smoke physics [episodes] [max_steps]
rl curriculum smoke robot [episodes] [max_steps]
rl curriculum smoke cybersecurity [episodes] [max_steps]
rl curriculum smoke healthcare [episodes] [max_steps]
rl curriculum smoke mathematics [episodes] [max_steps]
rl curriculum export [path]

rl domains [category]
rl domain <id>
rl sim game smoke [episodes] [max_steps]
rl sim physics smoke [episodes] [max_steps]
rl sim robot smoke [episodes] [max_steps]
rl sim energy smoke [episodes] [max_steps]
rl sim windows smoke [episodes] [max_steps]

rl net status
rl net modes
rl net enable <agent>
rl net mode <agent> <dqn|dueling-dqn|actor-critic|bandit|hierarchical|multihead-dueling|worldmodel|reward-model>
rl net save|load <agent> <path>
rl net reward <agent> good|bad
rl net wm|rm <agent> on|off

rl cartpole [episodes] [dueling|a2c|hierarchical|multihead] [shield] [tripwire] [curiosity]
rl life status|smoke|suggest   # alias of life rl …

rl real envs
rl real stage <agent> <sim|offline|shadow|gatedlive|live>

Domains

Run rl curriculum domains to print the live catalog. Run rl curriculum domain <id> to print actions, safety model, deployment stage, and a ready-to-run smoke command for a single domain.

Run rl curriculum smoke mathematics or nn math smoke mathematics to smoke all mathematics fields (can be long). Use nn math train <domain_id> for MathNN DQN training. Short aliases: pde, linalg, topology, ode, …

The generic smoke domains use a synthetic eight-value observation: progress, quality, risk, uncertainty, budget, compliance, load, time. Their actions are catalog-specific labels, but the smoke environment maps them into safe policy families: observe, optimize, explore, escalate, recover/fallback, and audit/deploy. These smoke tests do not touch external services.

Training Flow

1. Prove MathNN: nn demo, optional rl cartpole 50.
2. Run rl curriculum smoke all 8 96.
3. Check rl curriculum status and rl curriculum next.
4. Repeat smoke runs with more episodes for weak domains.
5. For math tutoring domains: nn math smoke mathematics / nn math train <id>.
6. For Life RL (advisory): life rl smoke all 8 then life rl suggest ….
7. Train real game/Minecraft policies only after simulator returns improve. Minecraft: dry-run → minecraft dqn net on → short train; optional minecraft dqn plan on for Executive setup steps.
8. For high-stakes domains such as healthcare, parenting, finance, legal, cybersecurity, infrastructure, and robotics, keep training synthetic or logged-data only until evaluation is strong and the safety model is reviewed.
9. Keep real-world domains in Sim or Offline until a checkpoint is promoted.
10. Move through Sim -> Offline -> Shadow -> GatedLive -> Live using rl real stage ... and the existing RLDeployment gates.
11. Save neural checkpoints: rl net save <agent> <path>.

Lua Helpers

ai_rl_status()
ai_rl_seed(n)
ai_rl_curriculum_status()
ai_rl_curriculum_plan()
ai_rl_domains()
ai_rl_domains("finance")
ai_rl_domain("healthcare")
ai_rl_curriculum_domains()
ai_rl_curriculum_domain("cybersecurity_soc")
ai_nn_modes()
ai_nn_policy_enable("life", "dueling-dqn", 16, 8)