[ElmoPA]

Adds an independent world-model path for joint visual-feature, dynamic XYZ-track and measured-pose prediction through three modality experts with shared attention. Embodiment-specific image residual adapters and pose interfaces support human, YAM, LIBERO gripper and Sharpa schemas.

The continuous recipe samples 2–16 recorded timestamps within ±3 physical seconds of an anchor; the discrete comparison uses four fixed-stride frames. Visual and track denoising warm up before pose loss ramps through the predicted context. Arbitrary temporal conditioning and selective prediction refinement support past, future and interpolation queries.

Prepared-data adapters enforce track availability semantics, DINO patch-grid/provenance checks, episode-level splits and train-only normalization. Includes native simulator pose/track preparation, resumable feature caching, isolated DINO diagnostics, and checkpoint-safe runner/RNG/sample-cursor recovery.

Validation: native simulation optimizer, validation and strict checkpoint-reload checks for gripper BC and both co-training recipes, plus focused model/data/runner tests. Real-world YAM acquisition clocks and dynamic metric tracks remain unresolved; this PR does not establish real-world training readiness. Historical run/checkpoint sources remain unchanged.

[ElmoPA]

• Add paired modality ablations for joint image forecast quality #265 : 2 dependent PRs (#276 , #320 )
• Train joint image, proprio and action MoT with smaller vector experts #264
• Add modular pixel video training and resilient log mirroring #262
• Optimize MoT attention and enable Flash4 BF16 defaults #261
• Add continuous-time MoT world model and offline data pipeline #260 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.