Under Review

AdaptManip: Learning Adaptive Whole-Body Object Lifting and Delivery with Online Recurrent State Estimation

A fully autonomous framework for humanoid robots to perform integrated navigation, object lifting, and delivery.

Morgan Byrd
Georgia Institute of Technology
Donghoon Baek
Georgia Institute of Technology
Kartik Garg
Georgia Institute of Technology
Hyunyoung Jung
Georgia Institute of Technology
Daesol Cho
Georgia Institute of Technology
Maks Sorokin
Georgia Institute of Technology
Robert Wright
Georgia Tech Research Institute
Sehoon Ha
Georgia Institute of Technology
Paper (PDF) arXiv Cite

Overview Video

Method

AdaptManip integrates three core capabilities into a unified autonomous pipeline: navigation to approach the target object, object lifting with adaptive grasp and manipulation control, and delivery to transport the object to a goal location. A recurrent pose estimator provides robust state estimation even under intermittent and noisy vision observations, enabling reliable closed-loop loco-manipulation.

Method Overview

Results

Object Shape Generalization

Cuboid — X-axis
Cuboid — Y-axis
Cuboid — Z-axis
Cylinder — X-axis
Cylinder — Y-axis
Sphere

Mass Variation

Light Mass (0.1x Nominal)
Heavy Mass (5x Nominal)
Mass Range

Adaptive Regrasp

Regrasp Example 1
Regrasp Example 2
Regrasp Example 3

Citation

@misc{byrd2026adaptmanip, title={AdaptManip: Learning Adaptive Whole-Body Object Lifting and Delivery with Online Recurrent State Estimation}, author={Morgan Byrd and Donghoon Baek and Kartik Garg and Hyunyoung Jung and Daesol Cho and Maks Sorokin and Robert Wright and Sehoon Ha}, year={2026}, eprint={2602.14363}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2602.14363}, }