OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction
arXiv:2509.26633v3 Announce Type: replace-cross Abstract: A dominant paradigm for teaching humanoid robots complex skills is to retarget human motions as kinematic references to train reinforcement learning (RL) policies. However, existing retargeting pipelines often struggle with the significant embodiment gap between humans and robots, producing physically implausible artifacts like foot-skating and penetration. More importantly, common retargeting methods neglect the rich human-object and human-environment interactions essential for expressive locomotion and loco-manipulation. To address th
This research addresses a critical limitation in current humanoid robot training methodologies by offering a more robust data generation technique, which is timely given the increasing focus on advanced robot capabilities.
Improving the fidelity and intelligence of human-robot interaction and loco-manipulation is a core challenge in robotics, and this development directly contributes to faster, more effective deployment of humanoid robots.
The ability to generate more realistic and interaction-preserving data accelerates the training of reinforcement learning policies for humanoid robots, making them more capable of complex tasks in diverse environments.
- · Humanoid robotics manufacturers
- · AI/ML research institutions
- · Automation sector
- · Logistics and manufacturing
- · Companies reliant on less sophisticated robotic solutions
More capable and agile humanoid robots are developed with greater efficiency.
This accelerates the adoption of humanoid robots in practical applications, potentially displacing human labor in certain physical tasks.
The enhanced realism in robot interaction training could lead to unexpected new applications and broader societal acceptance of advanced autonomous systems.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI