Phase 54.2 Spec & Technical Discussion: DynamicsPredictor

[web3guru888]

DynamicsPredictor — Spec & Technical Discussion

Technical discussion thread for Phase 54.2: DynamicsPredictor (#1035).

Core Design Questions

1. RSSM vs Transformer dynamics: DreamerV3 uses RSSM (GRU + stochastic); IRIS/TransDreamer use Transformers. Should we support both? What is the default?

2. Stochastic vs deterministic predictions: When is stochastic dynamics modeling essential (stochastic environments) vs. overkill (deterministic games)? Can we auto-detect?

3. Rollout stability: Multi-step rollouts accumulate error. What techniques should we use — scheduled sampling, noise injection, uncertainty-based truncation?

4. Action conditioning: How should actions be injected into the dynamics model — additive, multiplicative, FiLM conditioning, or cross-attention?

5. Temporal hierarchy: Should we implement multi-timescale dynamics (coarse abstract predictions + fine-grained motor)? This connects to hierarchical planning.

Proposed API Surface

class DynamicsPredictor:
    def predict_next(self, state: LatentState, action: Action) -> LatentState
    def rollout(self, state: LatentState, actions: List[Action]) -> List[LatentState]
    def get_uncertainty(self, state: LatentState, action: Action) -> float

Share your thoughts below.

Related: #1035 | #1033