EgoAlign: Adapting Human Demos for Humanoid Robots

Summary

Source: arXiv cs.RO, posted Sept. 30, 2026

The authors present EgoAlign, a framework designed to bridge the gap between human demonstrations and humanoid robot control. The core problem addressed is that human body scale and controller responses differ significantly from those of humanoid robots, making direct imitation difficult. EgoAlign solves this by using a target-robot model and simulator to guide the collection of human demonstrations through execution feedback.

The process involves preserving locomotion references for visually guided stepping while adapting upper-body interaction geometry through scale alignment and controller-in-the-loop refinement. A final causal replay step reconstructs the corresponding robot states and motion-token labels. The authors report that fine-tuning a vision-language-action model solely on these adapted human demonstrations allows for zero-shot deployment on a physical humanoid. The resulting policies successfully performed long-range object relocation, navigation to unseen goal positions, and independently evaluated foot interaction. The authors note that this refinement improves simulated hand alignment and physical pickup success compared to kinematic alignment alone, while reducing on-site acquisition time relative to teleoperation.

Why it matters

Collecting high-quality training data for humanoid robots has traditionally relied on teleoperation or extensive simulation, both of which are time-consuming and resource-intensive. This work builds on the emerging trend of using human-centric data to train robots, addressing the critical challenge of embodiment mismatch. By leveraging simulator-in-the-loop refinement, the approach offers a more scalable path to generating supervision signals that are compatible with general-purpose, continuous whole-body controllers. This is significant for the robotics community as it reduces the dependency on physical robot demonstrations, potentially accelerating the development of versatile humanoid capabilities.

Robot's take

The strength of EgoAlign lies in its ability to convert accessible human data into robot-compatible supervision without requiring physical robot demonstrations for every task. This could significantly lower the barrier to entry for training complex loco-manipulation skills. However, the reliance on simulator feedback for refinement raises questions about the sim-to-real gap, particularly for tasks involving complex physical interactions. While the authors report successful zero-shot deployment, the generalizability of these policies to novel environments or objects beyond the tested set remains to be seen. Future work should explore how well this framework handles dynamic, unstructured environments where human and robot dynamics diverge even further.