H2RB Builds a Bridge from Human Demonstrations to Robot VLAs

Introduction

For many robot-learning systems, the hardest part of deploying a pretrained policy is not model training itself but collecting demonstrations for the target robot. Humans can demonstrate manipulation skills quickly, yet their bodies, camera viewpoints, hand structures, and control signals differ substantially from those of robots. As a result, a human trajectory cannot simply be copied as a robot action sequence.

H2RB, or Human-To-Robot Bridge, addresses this gap while adapting a pretrained vision-language-action policy. The framework is designed to work with unpaired human and robot demonstrations, without requiring target-robot demonstrations for every task or synchronized human-robot trajectories. Its central idea is to build separate bridges for observations and actions rather than treating embodiment transfer as a direct motion-translation problem.

Key design choices

• Geometry-anchored representation alignment: H2RB uses hand geometry to place human and robot observations in a more compatible representation space. This encourages the policy to focus on spatial relationships among the hand, manipulated object, and target instead of overfitting to the identity of the acting body.
• Decoupled action decoding: A decoupled-decoder tokenizer allows the two domains to share action latents while retaining different decoding routes. Human demonstrations can therefore supervise behavior without requiring human and robot actions to obey one universal mapping.
• Heterogeneous co-training: Human and robot demonstrations are trained together, allowing human data to contribute behavioral information while the pretrained VLA retains its visual, semantic, and control priors.

What the reported results show

On real-world tasks, H2RB raises object-transfer performance from 8.0 to 51.8. In settings where co-training baselines fail to transfer skills demonstrated by humans, the method reaches a skill-transfer score of 62.7. The advantage also remains under object-placement changes: out-of-distribution transfer improves from 6.0 to 45.3.

These results point to an important distinction in cross-embodiment learning. Some parts of a demonstration—such as the intended behavior and the relationship between an object and a goal—can be shared. Other parts, including the exact motion commands, must remain embodiment-specific. H2RB is designed around that separation.

Significance and open questions

The approach suggests a scalable data strategy for robotics. Human demonstrations can provide relatively accessible behavioral supervision, while pretrained robot policies and robot-side data preserve execution constraints. This may reduce the need to gather extensive demonstrations separately for every task and robot configuration.

The available material does not provide the full task list, dataset sizes, robot configurations, or detailed ablations. It is therefore too early to determine how sensitive H2RB is to demonstration quality or how broadly it generalizes across robot forms. The paper nevertheless offers a clear direction: cross-embodiment transfer may benefit more from separating shared intent from embodiment-specific control than from seeking a single universal action translator.

Source: Hugging Face Daily Papers