VLCM checkpoints
Research checkpoints for multi-agent visual task-progress prediction and associated AO behavior cloning.
| File | Description |
|---|---|
cfm/Future5_epoch027_step152758.pt |
Frozen CFM-Future5 progress model, step 152,758. Inference-only export with exactly the original trainable tensors. |
cfm/config.json |
Forecast timing, base-model identifier and reconstruction metadata. |
bc/AO400_optimal_epoch051/bc_policy.pt |
Selected pure BC policy, epoch 51, trained on 320 optimal AO episodes and validated on 80 disjoint episodes. |
bc/AO400_optimal_epoch051/actor_bundle.json |
BC architecture, input/action contract and validation metadata. |
CHECKSUMS.json |
SHA-256 file identities and original CFM tensor identity. |
The CFM checkpoint requires the original custom VLCM model definition and the Qwen/Qwen3-VL-2B-Instruct base model. It is a partial custom-state checkpoint, not a standalone Transformers AutoModel. Restore compatible frozen custom parameters using reconstruction seed 42. The auxiliary success head is unused for progress scoring. This export excludes optimizer/scheduler/RNG state and private training configuration; the original resume checkpoint remains on the source cluster.
CFM-Future5 takes eight observations at 5 Hz and predicts ten progress bins at 0.2-second spacing. Its legacy action conditioning uses one command snapshot per bin, rather than every 30 Hz command. It predicts task progress, not collision probability or an RL value function. Current policy experiments use only the first +0.2-second forecast.
The associated BC policy is a small CNN with an eight-command prediction head and tanh command parameterization. It uses overhead RGB224, nonprivileged robot/goal/team-summary vectors and a fixed AO task ID. The selected validation first-action MAE is 0.0928846 in normalized command units; this is not a closed-loop success rate.
No datasets, credentials, authentication caches, or optimizer states are included in this repository. Refer to the original base-model terms for the Qwen dependency.