Go2 DimOS Replay Latent Dynamics Head

This is an experimental WorldForge-style world-model head trained on the derived WorldForge Go2 DimOS Replay World Pairs dataset.

The current package uses 2,557 action-conditioned current/future Go2 replay pairs from six usable public DimOS replay DBs.

What Was Trained

Only a small ridge dynamics head was trained:

frozen DINOv2 current-frame latent + egomotion/action delta
-> predicted residual future DINOv2 latent
-> current latent + residual = predicted future DINOv2 latent

The DINOv2 backbone remains frozen. This is not a trained V-JEPA model, not a Go2 foundation model, and not a safety-certified controller.

Evaluation

• Test future-latent cosine mean: 0.543147
• Test no-motion cosine baseline: 0.524954
• Test cosine lift vs no-motion: 0.018193
• Test candidate scoring accuracy: 0.258486
• Validation future-latent cosine mean: 0.631262
• Validation no-motion cosine baseline: 0.580600
• Validation cosine lift vs no-motion: 0.050662
• Validation candidate scoring accuracy: 0.284595

Candidate scoring uses the WorldForge-style contract:

score(candidate) =
  cosine(predicted_future_latent(current_image, candidate_delta), goal_future_latent)

For evaluation, the real future frame provides the goal latent and the real egomotion delta is ranked against counterfactual deltas.

Limitations

• Tiny replay-derived dataset.
• Egomotion labels come from pose deltas, not raw joystick commands.
• The model is intended as an inspectable research/demo artifact.
• It should not be used for direct robot control or safety decisions.