HLO-AMP: Hierarchical Loco-Manipulation from Object-Aware Motion Priors

SUMMARY

A paper on arXiv by Jean-Pierre Sleiman, Bilal Hammoud, He Li, Jonathan Foster, Sylvain Bertrand, Jessica K. Hodgins and Farbod Farshidian describes HLO-AMP, a hierarchical approach to humanoid loco-manipulation. It was submitted on 3 November 2026 and has been accepted at the Conference on Robot Learning (CoRL 2026).

The authors start from a common set-up in which a task-agnostic whole-body controller sits under a downstream task policy. They note that such controllers are often trained with manual reward shaping, which gives visibly machine-like motion, and without the manipulated object in the scene. Object dynamics are approximated with randomised force disturbances at the end-effectors. The paper says this proxy does not reproduce the multi-contact wrench that a real payload applies across the arms and torso, and that the gap becomes critical near the platform's payload limit.

Their alternative is to train the low-level controller as an object-aware motion prior. It is a single policy trained by motion imitation on retargeted motion-capture references covering the core motion primitives of the downstream task, with the object physically present in the scene. The paper says this stage inherits human-like coordination from the references and absorbs the contact wrench at its source, while keeping the compact task-space command interface of command-tracking controllers.

A reference-free high-level policy is trained on top. Its action channel is anchored to the empirical envelope of the low-level training data, so it can extend the priors to a long-horizon, multi-stage task while structurally avoiding queries that would push the low level off-distribution.

The framework is instantiated on a Unitree G1 for bimanual box pick-and-place. The authors report that the same deployed policy generalises across object masses and dimensions and across initial and goal poses, and that it recovers reactively from intentional drops during carrying.

Drafted with AI assistance from the source and reviewed by WBH. Follow the source link for the full text.