Cosmos3-Nano Bridge Forward-Dynamics World Model
Action-conditioned forward-dynamics (FD) world model for the BridgeData / WidowX setup,
post-trained from Cosmos3-Nano. Given a single conditioning frame plus a chunk of robot actions,
it predicts the resulting multi-view video — used as the rollout environment in a WorldGym-style policy
ranking over our bridge WM.
- Base model: Cosmos3-Nano (Qwen3-VL-8B MoT backbone + Wan2.2 VAE tokenizer)
- Mode:
forward_dynamics — condition on frame 0 + all actions → predict video
- Action space: bridge 7D
[Δxyz, Δrpy, gripper(abs)]
- Views: multi-view concat (3-camera WidowX)
- Chunk length: 16 · operating point: resolution 480, shift 5
- Precision: bf16
Checkpoints
Each subfolder is a self-contained HF safetensors export (7 shards + config.json + checkpoint.json).
| subfolder | iter | held-out motion-PSNR | PSNR | SSIM | note |
|---|
best-iter14000/ | 14,000 | 20.99 dB | 25.43 | 0.862 | best by held-out motion-PSNR |
latest-iter30400/ | 30,400 | 20.36 dB | 24.93 | 0.857 | latest; metric plateaued |
Held-out eval = seeded 1% split (never trained on), n=6 episodes, motion-PSNR masked to moving regions.
The single-chunk metric plateaus / slightly regresses past ~14k iters; best is the recommended default,
latest is provided for completeness.
Training
- g512: 16× H200 (2 nodes), global batch 512,
bridge_multimodal (WidowX, multi-view).
- Post-trained from the base Cosmos3-Nano DCP checkpoint via the
cosmos-framework action SFT recipe.
- Note: the bridge multi-node dataloader occasionally stalled on NFS during training (recovered by resubmit).
Usage
These are cosmos-framework model exports (not transformers-native). Load with the cosmos-framework
inference stack (the exported config.json / checkpoint.json describe the OmniMoTModel). Point the
bridge multiview FD rollout tooling at a subfolder as the checkpoint path.
Caveats
Research checkpoints. Bridge episodes are short (~≤98 frames); FD conditions on a single history frame so
long autoregressive rollouts drift. Best for short-horizon dreaming and policy-ranking use.
