Exact equivariance, kept through training, buys zero-shot generalisation across the symmetry group
TLDR
An equivariant latent world model achieves zero-shot generalization across symmetry groups by keeping prediction loss invariant, verified with real-world experiments.
Reasoning
The paper provides strong theoretical guarantees and empirical validation for equivariant world models, showing significant improvements over non-equivariant baselines. However, the abstract lacks details on scalability and broader applicability beyond the tested symmetries.
Read-first score
Read-first score 62, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.
Field roles
Rank sensitivity
Stability: volatile; rank range: 433.
Keyword Scores
Deep Analysis
Innovations
- Exact equivariance kept through training buys zero-shot generalization across the symmetry group
- Provable symmetry of training loss: one-step prediction relMSE exactly invariant across group, so fitting on restricted slice determines entire orbit
- Verification at laptop scale (CPU/MPS) with full seeding, showing symmetry survives real optimizers (Muon/AdamW + EMA + VICReg) with residual ~1e-6
- One-step error flat to five digits across group while non-equivariant baseline breaks OOD (VN ×1.00 vs baseline ×13.8 in 2D, ×17.2 in 3D, ×157 over full SE(3)), with equivariant model 4.5-7.4× smaller
- Closed-loop: under matching equivariant planner, control trajectory at orientation g is exactly ρ(g) applied to seen one, closed-loop error invariant across group (float-floor-exact in 2D/SO(2) on PushT, statistically flat in 3D/SE(3) with disjoint 95% CIs)
- Stress-test against Sutton's Bitter Lesson: augmentation, brute-force scale, and soft-equivariance each close at most the across-group task metric, never the float-floor exactness
- H-fold rollout stays flat (×1.00, ≤2×10⁻⁷) at every horizon while baseline residual compounds with H
Methodology
A latent world model is built from an equivariant encoder E and an equivariant predictor f, with a group G acting on latents via orthogonal representation ρ(g). The training loss is one-step prediction relMSE, which is exactly invariant across the group, so fitting dynamics on a restricted slice of orientations mathematically determines it on the entire orbit. The model is trained end-to-end at laptop scale (CPU/MPS) with full seeding, using Muon/AdamW + EMA + VICReg, and compared against a non-equivariant baseline of the same hypothesis class. Evaluation metrics include one-step error, closed-loop error, and rollout error across the group.
Key Results
The equivariant model achieves one-step error flat to five digits across the group, while the non-equivariant baseline exhibits out-of-distribution error factors of 13.8× (2D), 17.2× (3D), and 157× (full SE(3)), despite the equivariant model being 4.5-7.4× smaller. Closed-loop error is invariant across the group (float-floor-exact in 2D/SO(2) on PushT, statistically flat in 3D/SE(3) with disjoint 95% CIs), and H-fold rollout remains flat (×1.00, ≤2×10⁻⁷) at every horizon, whereas the baseline residual compounds with H.
Limitations
- Task-success sweeps are out of scope
- Planner-free invariance is out of scope
- Scaling is out of scope