CASE STUDY
How Roborax delivered a cross-embodiment training dataset that cut sample requirements by 3x for a humanoid robot foundation model team.
A leading humanoid robotics team was building a general-purpose foundation model and needed a training dataset that spanned multiple task types, multiple environments, and multiple robot embodiments. Their internal data collection capability could not produce data at the breadth and pace the model required.
Foundation models for humanoid robots require training data that generalises across a wide distribution of tasks and environments. Single-embodiment, single-environment datasets produce policies that overfit. The team needed data from at least three distinct robot platforms, across six task categories, collected in five different environment types — a scope far beyond what their internal team could execute.
Roborax deployed dedicated operator teams across three humanoid platforms simultaneously. Over a 12-week program, we collected 2.4 million trajectories spanning whole-body manipulation, loco-manipulation, bimanual coordination, and unstructured navigation. Data was delivered in RLDS format with per-trajectory quality scores and failure mode labels.
The client’s foundation model, trained on the Roborax dataset, required 3x fewer demonstration samples to reach the same policy performance as a model trained on their prior single-embodiment dataset. Sample efficiency improved because the training distribution was genuinely diverse — not just larger.
Related: Humanoid robotics solutions — Case studies.
For humanoid robot foundation model teams, sample efficiency is a direct commercial constraint. Generating training data is expensive — every additional demonstration episode has a real cost. A 3x improvement in sample efficiency means the same policy performance can be achieved at one-third of the data cost, or that the data budget can be applied to three times as many task categories. At foundation model scale, this difference is commercially significant. Related: Humanoid robotics solutions — Case studies.
