Train world models in N-dimensional physics, then transfer policies back to 3D tasks with reproducible, staged validation.

Core Idea
├─ Learn in flexible N-D simulation
├─ Transfer structure across dimensions
└─ Validate robustness with reproducible seed-based reports

• 🎯 Build a research-grade playground for world model + transfer experiments.
• 🧪 Keep everything measurable via multi-seed summaries and significance reports.
• 🚚 Run locally or on Kaggle with one operational path (no Colab dependency).

[Physics Engine (N-D)]
        |
        v
[PushBallNDEnv]
        |
        v
[World Model]
  |---- MLP
  |---- GRU
  |---- RSSM
  |---- Physics+Residual
        |
        v
[DreamTrainer + ReplayBuffer]
        |
        v
[Experiments]
  |---- baseline
  |---- transfer
  |---- ablation
  |---- robustness
  |---- curriculum
  |---- hifi_migration

Roadmap
├─ P0 Baseline Freeze: ✅
├─ P1 Semantic Interface Lock: ✅
├─ P2 Domain Randomization: ✅
├─ P3 Physics+Residual Model: ✅
├─ P4 Curriculum Runner: ✅
├─ P5 Aggregate Reporting: ✅
├─ P6 Hi-Fi Proxy Migration: ✅
└─ De-Colab + Kaggle Consistency: ✅

Source: report/release_significance_p0_vs_p2v2_5seed.md

Source: results/p0_freeze/p0_freeze_9seed/p0_summary.json

pip install -r requirements.txt
pytest -q
python experiments/run_baseline.py

python experiments/run_p0_baseline_freeze.py \
  --run-id-prefix p0_freeze_5seed \
  --seeds 11 22 33 44 55 \
  --baseline-epochs 8 \
  --transfer-pretrain-epochs 6 \
  --transfer-finetune-epochs 6 \
  --robustness-episodes 120

python experiments/run_p0_baseline_freeze.py \
  --run-id-prefix p2_hard020_5seed \
  --seeds 11 22 33 44 55 \
  --domain-rand \
  --domain-rand-scope robustness_only \
  --robustness-domain-rand-difficulties hard_only \
  --domain-rand-scale 0.20 \
  --domain-rand-profile conservative \
  --domain-rand-warmup-episodes 200 \
  --domain-rand-warmup-epochs 8 \
  --domain-rand-source-multiplier 1.0 \
  --domain-rand-finetune-multiplier 0.5 \
  --baseline-epochs 8 \
  --transfer-pretrain-epochs 6 \
  --transfer-finetune-epochs 6 \
  --robustness-episodes 120

python experiments/significance_report.py \
  --a-prefix p0_freeze_5seed \
  --b-prefix p2_hard020_5seed \
  --report-name release_significance_p0_vs_p2hard020_5seed

Runtime Policy
├─ Local + Kaggle only
└─ Colab path removed from active pipeline

python kaggle_job_manager.py run \
  --owner <your_kaggle_username> \
  --slug high-dimensional-worldmodel-aggressive \
  --seed 11

• Direct stage execution inside kernel: baseline/transfer/ablation/robustness.
• run_id and seed are verified to propagate end-to-end.

.
├─ envs/
├─ physics/
├─ models/
├─ training/
├─ experiments/
├─ kaggle/
├─ report/
├─ results/
└─ RUNBOOK.md

Higher-dimensional world models for cross-dimensional transfer: reproducible N-D physics training, robustness evaluation, and Kaggle-ready experiment orchestration.

(optional) link to your latest report or project page

reinforcement-learning, world-model, model-based-rl, transfer-learning, simulation, physics, pytorch, kaggle, reproducibility, research

• RUNBOOK.md for execution commands.
• report/ for technical and significance reports.
• 改造计划.MD for phased roadmap (P0 -> P6).
• Latest smoke validation reports:
  • report/kaggle_smoke_hard020_2seed_summary.md
  • report/kaggle_smoke_vs_hifinal_hard020_overlap2_significance.md