---
license: apache-2.0
tags:
- world-model
- active-learning
- confidence-estimation
- robot-learning
---

# ConfAL-WM · Model Checkpoints

Model weights for **ConfAL-WM: Confidence-Guided Active Learning for Action-Conditioned
World Models** (anonymous submission). These artifacts correspond to the
"07 · Models & Data" section of the project page.

All checkpoints were produced inside the release codebase; absolute paths and
machine-specific metadata have been scrubbed (`<DATA_ROOT>` / `<ANON…>` placeholders).

## Checkpoints

| File | Card | Description |
|---|---|---|
| `EVAC warmup v1.ckpt` | EVAC · Warmup v1 | RoboTwin2.0 domain-adapted warmup world model. Starting point for all active-learning rounds (v1 inference + confidence probe scoring). Lightning ckpt, epoch 10 / step 2000. |
| `EVAC-v2 weighting none.ckpt` | EVAC-v2 · Weighting None | Selection-only retrained checkpoint (mean-risk acquisition, seed 123). Lightning ckpt, epoch 8 / step 4000. |
| `EVAC-v2 weighting frame.ckpt` | EVAC-v2 · Frame | Confidence-guided **frame**-level weighting (mean-risk acquisition, seed 42). Lightning ckpt, epoch 8 / step 4000. |
| `EVAC-v2 weighting frame+patch.ckpt` | EVAC-v2 · Frame + Patch | Dense confidence-guided (frame + patch) weighting (mean-risk acquisition, seed 3407). Lightning ckpt, epoch 8 / step 4000. |
| `Confidence probe RoboTwin2.0.pt` | Confidence Probe · RoboTwin2.0 | Main C3 confidence probe used in the paper (probe step 6000). Takes EVAC decoder features (h_dec embeddings) and outputs per-frame/patch confidence. |
| `Confidence probe AgiBotWorld.pt` | Confidence Probe · AgiBot World | Additional confidence probe trained on AgiBot World (probe step 6000). |
| `YOLO RoboTwin2.0.pt` | YOLO · RoboTwin2.0 | Gripper/trajectory-metric detector (left/right gripper) for EWMBench-style evaluation. Ultralytics format; train args sanitized. |

## Usage notes

- The `EVAC*` checkpoints are PyTorch-Lightning archives; restore with
  `LightningModule.load_from_checkpoint(...)` using the model definition in the code release.
- The probes are plain `torch.save` state dicts; load with `torch.load(..., map_location="cpu")`.
- The YOLO detector can be loaded directly with `ultralytics.YOLO(path)`.
- Companion precomputed data (v1 inference outputs, dense confidence maps, baseline
  scoring artifacts, YOLO annotations, evaluation tables) is available in the dataset repo
  `anonymous89793/ConfAL-WM-Dataset`.

## Anonymization

- All absolute filesystem paths inside metadata/pickles were replaced with placeholders
  (`<DATA_ROOT>/`, `<ANON…>`); no usernames, hostnames, or machine paths remain.
- Checkpoint tensors were **not** modified — only metadata strings.