tags:
- robotics
- thinkwam
- libero
- robotwin
- robocasa
ThinkWAM checkpoints
ThinkWAM inference checkpoints exported from the selected training checkpoints.
Each .pt file contains model, config, and step. Optimizer, scheduler, and random-number-generator states are excluded.
Every exported model tensor was compared byte for byte against its source checkpoint, including its dtype and shape. The original BF16/FP32 storage precision and model configuration are preserved.
| Benchmark | Checkpoint | Stored step |
|---|---|---|
| LIBERO / LIBERO-Plus | libero_thinkwam_checkpoint.pt | 75000 |
| RoboTwin 2.0 | robotwin2.0_thinkwam_checkpoint.pt | 155000 |
| RoboCasa24 | robocasa24_thinkwam_checkpoint.pt | 20000 |
The LIBERO checkpoint is shared by LIBERO and LIBERO-Plus evaluations. The RoboCasa24 step is the counter saved in its continuation run, not a claim about total training steps across runs.
Loading these checkpoints requires the compatible ThinkWAM implementation. Configurations are embedded in each checkpoint and are also provided as separate JSON files for inspection. Runtime paths and benchmark resources must be configured for the evaluation environment.
File sizes and SHA-256 checksums are recorded in checkpoint_manifest.json and SHA256SUMS.