Skip to main content

Unified Training Comparison

This smoke demo shows that DreamerV3 and TD-MPC2 can run through the same high-level training flow and the same quick verification flow.

It uses the same random ReplayBuffer source for both families, so treat it as a contract demonstration rather than a benchmark or a real-environment performance comparison.

Run​

uv sync --extra dev --extra training
uv run python examples/compare_unified_training.py --quick

What It Demonstrates​

  • same random ReplayBuffer source for both families
  • same TrainingConfig contract
  • same quick verification flow for both families
  • same artifact generation flow
  • same visualization helper (write_reward_heatmap_ppm)

The default command also runs quick verification for each family and writes the structured result alongside the training outputs. Use --skip-verify only when you intentionally want a faster local smoke without verification artifacts.

Artifacts​

  • outputs/comparison/summary.json
  • outputs/comparison/dreamer.ppm
  • outputs/comparison/tdmpc2.ppm
  • outputs/comparison/dreamerv3/quick_verify.json
  • outputs/comparison/tdmpc2/quick_verify.json

Interpretation​

This demo proves that both families can be created, trained, rolled out, and checked through the same public workflow shape. It remains a smoke-only contract demonstration and does not establish real-environment performance, benchmark superiority, paper reproduction, or a public proof claim.