These studies range from controlled robot experiments to document attribution and generated garment video. Their different evaluation designs make a single readiness label unhelpful.
The findings below are reported by the researchers in the cited studies. The evaluation questions are practitioner interpretation, not measured Hyperion client outcomes.
PhysBrain 1.0
PhysBrain 1.0 turns human-view video into structured supervision and transfers those priors into robot policies. The paper includes simulation and controlled tabletop robot experiments. These are meaningful task results, but human and robot embodiments differ and perception errors can propagate; broad enterprise deployment remains a separate evaluation. Research source.
Evaluation question: Test transfer on the target robot, objects and visibility conditions, including cases where annotation or depth estimates are wrong.
MMSkills
MMSkills packages procedures, runtime state and visual keyframes into skills that an agent can inspect when relevant. The authors report improvements on GUI and game benchmarks. Coverage depends on the source trajectories, errors can propagate from generated skills, and the extra inspection introduces overhead. Research source.
Evaluation question: Evaluate skill selection and recovery when the current interface differs from the demonstrations used to build the skill.
CiteVQA
CiteVQA requires a visual-document answer together with element-level evidence and scores the two jointly. Its experiments expose a gap between producing an answer and attributing it correctly, with marked differences among evaluated models. This benchmark helps test attribution; it does not establish that every cited passage is authoritative for a business decision. Research source.
Evaluation question: Require both the answer and the exact supporting document region, then independently check the support.
DexJoCo
DexJoCo provides simulation tasks and a toolkit for dexterous manipulation, including coordination, tool use and longer task sequences. The study compares training and adaptation choices within that environment. Simulation performance does not establish transfer to physical equipment, particularly where contact forces and tactile information matter. Research source.
Evaluation question: Separate success in simulation from acceptance on the intended hardware and contact conditions.
FashionChameleon
FashionChameleon studies coherent video generation with interactive switching between garments. The paper reports 23.8 frames per second at 720p on an NVIDIA H200 GPU. That configuration-specific result does not establish affordable live-commerce operation, broad garment coverage or an effect on returns or revenue. Research source.
Evaluation question: Measure appearance fidelity, switching latency and sustained cost on the intended hardware and garment range.
Match the test to the intended use
For robot transfer, test the actual embodiment and contact conditions. For documents, require supporting regions as well as answers. For video, record the hardware and sustained behavior needed to meet the interaction requirement.
