Open-HumanWAM
Collection
Model checkpoints for Open-HumanWAM. • 5 items • Updated • 4
HumanWAM-Vanilla-Exploration
RoboTwin post-trained HumanWAM-Vanilla checkpoints for the exploration studies in Open-HumanScale. This repository covers IA scaling, IDM scaling, matched diversity, cross-source alignment, demonstration efficiency, human action supervision, robot versus egocentric pretraining, and multi-view adaptation.
Checkpoints
Weights are organized under scaling-ia/, scaling-idm/, matched-diversity/,
cross-source-alignment/, demonstration-efficiency/, action-supervision/,
robot-vs-ego/, and multiview/ as each verified bundle becomes available.
Only RoboTwin post-trained weights are included.
Filenames identify the experimental setting and the study results that use the
checkpoint. For example, scaling-ia/ego-100h-clean-scene-ood.pt is the checkpoint
selected for the Clean and Scene-OOD results of IA scaling at 100 hours;
scaling-ia/ego-100h-random.pt is selected for Random. These names do not restrict
where a checkpoint can be evaluated and do not imply that all three best scores
in an experiment row come from a single checkpoint. point-NN follows increasing
x-axis order for cross-source and action-supervision learning curves.
Wrist alignment
Because the exploration studies span a long experimental period, some runs used different wrist-alignment conventions during RoboTwin post-training; for the relative-action comparisons, this difference affects only the absolute-state input, and we observed no noticeable performance difference across repeated comparison experiments. Absolute-action branches retain their own action/state transforms and matching statistics, as recorded in each checkpoint configuration.
Usage
Consult manifest.json for published weight paths and their matching files under
configs/, stats/, and evaluations/. Download the weight and exactly those
referenced files together. Statistics are specific to the action/state convention,
view layout, and demonstration subset; they are not interchangeable across series.
The common Wan dependencies and fixed revisions are recorded in the manifest.
Release status
Upload in progress. manifest.json distinguishes published weights from
pending entries. A pending entry is not an available download. The original .pt
files are preserved without precision conversion or resaving.
Exploration inference adapters are pending. Configuration/normalizer checks and successful uploads do not constitute an end-to-end simulator evaluation. The currently validated main-policy W0 adapter is not a general Exploration adapter and does not implement Scene-OOD. Per-series instruction, seed/shard, and execution details that remain unverified are marked pending rather than inherited from the main model. No new aggregate benchmark results are claimed.