Skip to content

Latest commit

 

History

History
53 lines (43 loc) · 4.36 KB

File metadata and controls

53 lines (43 loc) · 4.36 KB

Atlas.WM v4.0 - Results Ledger

Every claim in this project is backed by a committed script, a reproducible number, and (where applicable) a regression test. This ledger consolidates them. Dates: v4.0 rebuild executed 2026-07-02 → 2026-07-04 (PRs #25–#44).

Headline results

# Claim Evidence Where
1 Physics is identifiable from position-only observations under a random policy Closed-form median-of-ratios estimator: friction_agent R² = 0.865, MAE 0.006 on the noisy (σ=0.05) on-disk dataset scripts/oracle_friction_agent.py, guarded by tests/test_oracle_friction.py
2 A learned model identifies physics - given the right features Belief v2 (engineered features, window 40): gravity +0.43, friction_agent +0.22, friction_box +0.09 (raw-GRU baseline: −0.05/−0.49/−0.10) atlas_wm/data/dynamics_features.py, scripts/train_physics_belief.py
3 Data collection matters as much as the model Identical model+training, active vs random data: mean belief R² 0.246 → 0.341 (+39%); gravity 0.67 atlas_wm/data/exploration.py, regression-locked in tests/test_active_exploration.py
4 The AD-2 architectural guarantee is real z_static_immutable max drift over open-loop self-fed rollouts: exactly 0.0, incl. on MuJoCo scripts/evaluate.py, tests/test_rollout_training.py
5 The self-predictive objective is stable without scale crutches VICReg recipe: z-std 0.99, next-frame MSE 0.001750 → 0.000996 (−43%); frame stacking flipped from harmful to helpful atlas_wm/training/objectives.py, tests/test_objectives.py
6 Rollout training changes rollout behavior One-step wins h=1 (0.000993 vs 0.001157); rollout-trained wins h≥3 (h=10: 0.0357 vs 0.0389) training.rollout_k, scripts/evaluate.py
7 Physical structure generalizes where MLPs interpolate Symplectic-dissipative head: OOD gravity degradation +12% vs +69% (residual) - but absolute error 3.5× worse in-dist (honest negative) model.dynamics_head, tests/test_hamiltonian_head.py
8 The pipeline is environment-general Unchanged pipeline on real MuJoCo contact physics: h=1 MSE 0.001089 vs linear ceiling 0.000465 (2.3× gap vs 3.6× on the gridworld) atlas_wm/environments/mujoco_pointmass.py

| 9 | The identification thesis transfers to real contact physics (v4.1) | COAST/PUSH policy + physics-informed features on MuJoCo: learned belief friction +0.28, mass +0.11 (random + generic stats: R² ≈ 0); gravity alone structurally unidentifiable (only μ·g enters box dynamics) | atlas_wm/data/mujoco_features.py, CoastPushPolicy |

Retractions (v3.x claims that did not survive re-verification)

  1. "friction_agent is not identifiable" - wrong. The original oracle was never committed and its MSE objective is destroyed by bounce outliers; the environment also silently voided the box-physics signal (boxes exited the grid, fixed in B1). See MODEL_CARD "RETRACTED" section.
  2. "Belief R² 0–0.15 is a data-regime ceiling" - wrong. Numbers were confounded by a 20× data-scale mismatch between pipelines (fixed in B2) and probe label leakage from overlapping windows (fixed in B5).

Named findings

  • Anti-fragility to unobservables (B11): a deterministic info-seeking policy that dashed along a fixed line repeatedly struck unobservable obstacles and poisoned episode medians (heavy-tail R² collapse with near-unchanged MAE). Random walks don't repeat their mistakes; deterministic policies do. Golden-angle rosette rotation fixes it.
  • Evidence-length bound (B10): friction_agent's target std (0.025) is below the optimal estimator's window-level error at 18 steps (0.043) - SNR < 1 means no learner can score positive there; 38-step windows reach signal scale. Belief quality is bounded by evidence, and active exploration raises the per-step information rate.
  • Prediction grounding (B7): nothing in a latent-matching + current-frame-reconstruction loss trains the decoder∘dynamics composition inference actually uses; adding it cut ~20% error on top of the stable objective.

Known limits (also in MODEL_CARD)

Toy/point-mass environments only; no real-robot or visual claims; heuristic (not learned) exploration policy; process-noise random walk floors long-horizon error; the symplectic head underfits at current capacity.