aiqbench analysis
This paper presents FetchMan, a humanoid robot policy trained via sim-to-real transfer using reinforcement learning and Flow-GRPO to overcome the performance limits of behavior cloning in visual loco-manipulation tasks. By leveraging over 150,000 simulated scenes, the model achieves zero-shot deployment on a Unitree G1 with significant success rates for navigating unseen environments to grasp objects. The work marks an important step toward scalable generalist policies that can handle complex walking and manipulation simultaneously without extensive real-world data collection.
FetchMan: Learning Visual Humanoid Loco-Manipulation Policies from Simulated Experiences
arXiv:2608.17027v1 Announce Type: new Abstract: Visual loco-manipulation policies that can generalize to novel scenes and objects have long been a goal of robotics research. However, today's data-hungry algorithms make collecting sufficient demonstrations a struggle for tabletop manipulation, and even more so for humanoids that must also walk and balance. Learning from simulated data and transferring that behavior to the real world, as is commonly done in locomotion, sidesteps this struggle, so we replicate that recipe for loco-manipulation. In doing so, we find that cloning synthetic demonstrations results in a low performance ceiling no matter the amount of training data. Reinforcement learning breaks through it, and refining the cloned policy with Flow-GRPO on a single sparse reward yields performance that synthetic behavior cloning cannot match. Together, these stages form our end-to-end sim-to-real pipeline spanning more than 150,000 scenes, which we use to train FetchMan. We evaluate it on FetchMan-Bench, a simulation benchmark we release, and deploy it zero-shot on a real Unitree G1, where our single-object reach-and-pick policy walks to and grasps a target across unseen scenes at 73.3% success. Finally, we extend this recipe to multi-object training, a first step toward loco-manipulation generalist policies at this data scale.
Abstract:Visual loco-manipulation policies that can generalize to novel scenes and objects have long been a goal of robotics research. However, today's data-hungry algorithms make collecting sufficient demonstrations a struggle for tabletop manipulation, and even more so for humanoids that must also walk and balance. Learning from simulated data and transferring that behavior to the real world, as is commonly done in locomotion, sidesteps this struggle, so we replicate that recipe for loco-manipulation. In doing so, we find that cloning synthetic demonstrations results in a low performance ceiling no matter the amount of training data. Reinforcement learning breaks through it, and refining the cloned policy with Flow-GRPO on a single sparse reward yields performance that synthetic behavior cloning cannot match. Together, these stages form our end-to-end sim-to-real pipeline spanning more than 150,000 scenes, which we use to train FetchMan. We evaluate it on FetchMan-Bench, a simulation benchmark we release, and deploy it zero-shot on a real Unitree G1, where our single-object reach-and-pick policy walks to and grasps a target across unseen scenes at 73.3% success. Finally, we extend this recipe to multi-object training, a first step toward loco-manipulation generalist policies at this data scale.
Robotics (cs.RO)
arXiv:2608.17027 [cs.RO]
(or arXiv:2608.17027v1 [cs.RO] for this version)
https://doi.org/10.48550/arXiv.2608.17027 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 286 words
Related coverage
DreamMimic: Learning Visuomotor Whole-Body Loco-Manipulation via World Model
arXiv · Aug 25, 2026
MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation
arXiv · Jun 9, 2026
$\omega$-0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation
arXiv · Aug 7, 2026
Morphology-Aware Human Motion Retargeting for Wheeled-Humanoid Loco-Manipulation
arXiv · Sep 11, 2026