Semantic Latent Diffusion Policy
This repository stores checkpoint artifacts for the semantic latent action space experiments.
Scope
The uploaded artifacts cover the full BRIDGE experiment suite under full_bridge_seed0:
- RoLD baseline LAT.
- Multi-delta semantic LAT.
- Plain monolithic LDP on RoLD LAT.
- Plain monolithic LDP on multi-delta LAT.
- Semantic-intent LDP on multi-delta LAT.
- Semantic-intent + semantic-denoising LDP.
- K=4 RoLD and multi-delta expert policies.
- K=4 observation routers and semantic-intent router.
- Evaluation metrics, training configs, and launch scripts.
The BRIDGE dataset and cached R3M feature directories are not included.
Important Conditioning Note
The full dataset metadata contains task_name strings, but the completed runs used
task_label_mode=original, i.e. categorical task-id conditioning. These checkpoints
are not language-token-conditioned policies.
Main Full-Dataset Results
| Model | Test MSE | Motion MSE | Gripper MSE | Gripper Acc |
|---|---|---|---|---|
| RoLD plain LDP | 0.037981 | 0.001409 | 0.257416 | 0.7278 |
| Multi-delta plain LDP | 0.038589 | 0.001408 | 0.261680 | 0.7257 |
| Multi-delta semantic-intent LDP | 0.036052 | 0.001372 | 0.244134 | 0.7434 |
| Multi-delta semantic-intent + semantic denoising | 0.036084 | 0.001366 | 0.244389 | 0.7433 |
Key Interpretation
The multi-delta semantic LAT strongly improves visual-effect latent geometry, but a plain monolithic LDP does not automatically benefit. Semantic-intent LDP improves offline action prediction, supporting the representation-policy gap hypothesis.
Structure
checkpoints/full_bridge_seed0/: all full-dataset checkpoints, configs, metrics, and summaries.scripts/: launch/evaluation scripts used to run the experiments.MODEL_INDEX.md: map from experiment names to checkpoint paths.manifest.json: uploaded file manifest.
Repository target used by uploader: bageldotcom/semantic-latent-diffusion-policy.