Robotics · cross-embodiment mixability

crossembodiment-align

A manifest-first assessment of whether robot demonstration datasets are compatible enough to mix, with every recommendation tied to metadata and measured statistics.

Public · synthetic demo

The problem

Missing embodiment metadata can turn “more data” into an invalid mixture.

The audit checks control-rate, gripper, action-space, and metadata compatibility before training compute is spent.

The money shot

Two of three pairs are explicitly do-not-mix

Pairwise recommendations

pair	score	recommendation
bridge_v3_7dof ↔ franka_7dof	0.5725	do-not-mix
bridge_v3_7dof ↔ so101_6dof	0.3175	do-not-mix
franka_7dof ↔ so101_6dof	0.7450	mix-subset

Manifest audit

3embodiments

1manifest validation flag

2metadata flags

5 Hztarget control rate

The public BridgeData v3 conversion is known to leave robot_type empty; this is the kind of omission the manifest audit surfaces.

Honest limitations

The tool reads metadata and statistics only. It does not train a model, run a robot or policy, or verify that mixing improves actual performance.

This page reports the supplied sample output and its stated limits. Inspect the code and evidence ↗