Paper Overview
- Field: Machine Learning
- Authors: Xun Tu, Minghao Zou, Karthik Desingh
- Published: 2026-11-06
- arXiv: 2611.09864
- Model-based methods: often require accurate dynamics models.
- Learning-based methods: reduce this requirement, but typically assume a fixed or predefined initial grasp.
- Rearrangement success rate improved from 79.87% (best single-grasp policy) to 89.84% — a gain of 9.97 percentage points.
- Final pose error was reduced across all evaluated objects.
- Paper: https://arxiv.org/abs/2611.09864
Summary
Loco-manipulation — coordinating locomotion and manipulation over long horizons — is a key capability for large-object rearrangement in applications such as service robotics and warehouse automation. Reliable execution remains difficult because performance depends on complex robot-object and object-environment interaction dynamics.
Existing approaches fall into two camps:
However, different geometrically stable grasps can lead to substantially different downstream manipulation performance.
Method
The authors propose Grasp-TAGS, a task-aware grasp selection framework for goal-conditioned large-object rearrangement:
1. Train a unified PPO reinforcement learning policy across a variety of initial grasp configurations. 2. Distill the learned critic into a lightweight grasp utility network that ranks candidate grasps using only candidate grasp configurations and goal information. 3. At inference time, select the grasp with the highest predicted utility for the downstream policy — before contact occurs.