HumanWAM-Vanilla-RoboTwin

HumanWAM-Vanilla post-trained on RoboTwin 2.0, using 50 clean demonstrations per task across 50 tasks (2,500 training demonstrations). The policy predicts a 32-step action chunk and executes 24 steps before observing again.

Checkpoint

Item	Value
File	`model.pt`, PyTorch checkpoint
Selected post-training step	21,000
Post-training schedule	24,000 steps, A2 augmentation
Initialization	HumanWAM-Vanilla-Base: 50,000-hour pretraining at step 40,000, followed by EgoPro W0 mid-training at step 10,000
Sampling	10 video denoising steps followed by 10 action denoising steps
Action chunk	32 predicted steps, 24 executed steps per query

Reported results

The archived evaluation covers 50 tasks × 50 expert-valid episodes per condition, with unseen instructions for both Clean and Randomized: 5,000 episodes in total. Both results use the same released checkpoint.

Clean success (%)	Randomized success (%)	Mean (%)
92.08 (2,302 / 2,500)	68.88 (1,722 / 2,500)	80.48

These are archived results, not a new evaluation of this release. The manuscript version reviewed for this upload describes 100 episodes per task, while the archived score record contains 50. Reconciliation with the original simulator logs remains pending; the table above follows the archived record. This protocol differs from the 100-episode-per-task Clean-seen / Randomized-unseen protocol reported for HumanWAM-Goal-RoboTwin.

Usage

hf download DAGroup-PKU/HumanWAM-Vanilla-RoboTwin --local-dir HumanWAM-Vanilla-RoboTwin

The repository includes model.pt, config.yaml, dataset_stats.json, evaluation.json, and manifest.json. Use the matching W0 three-view adapter and common Wan resources. The validated adapter still depends on local code awaiting a pinned public release; see manifest.json for compatibility and validation status. No upstream pretraining checkpoint or training corpus is required for inference.

Action convention

Absolute poses in the camera frame, with native RoboTwin wrist orientation (unalign): no additional wrist alignment is applied. Use the policy's matching world-to-camera transform and normalization statistics. The wrist-alignment transforms and statistics from the aligned head-only scaling experiments are not interchangeable with this checkpoint.

Release status

The selected weight and its inference configuration, matching normalization statistics, and archived evaluation protocol are available. These files are bound in manifest.json. A single Clean episode has been completed using anonymously downloaded release files and the local W0 adapter; see demo_receipt.json. A pinned public adapter release remains pending. This demo does not constitute a new benchmark evaluation.