CageCoOpt: Enhancing Manipulation Robustness through Caging-Guided Morphology and Policy Co-Optimization

TL;DR

This work tackles robustness in non-prehensile robotic manipulation under object-shape uncertainty and unmodeled dynamics by proposing CageCoOpt, a hierarchical framework that co-optimizes manipulator morphology and control policy. The lower level learns a universal policy via PPO conditioned on morphology and object shape, while the upper level uses multi-task Bayesian optimization with a GP surrogate to select morphologies, guided by the Minimum Escape Energy (MEE) metric. By integrating MEE into both optimization levels, the approach promotes energy-bounded caging configurations that improve success under uncertainty and transfer to real-world hardware. Four manipulation tasks in simulation and real experiments demonstrate that co-optimization yields more robust and reliable manipulation compared to unoptimized baselines, highlighting caging as a viable strategy for robust non-prehensile control.

Abstract

Uncertainties in contact dynamics and object geometry remain significant barriers to robust robotic manipulation. Caging mitigates these uncertainties by constraining an object's mobility without requiring precise contact modeling. However, existing caging research has largely treated morphology and policy optimization as separate problems, overlooking their inherent synergy. In this paper, we introduce CageCoOpt, a hierarchical framework that jointly optimizes manipulator morphology and control policy for robust manipulation. The framework employs reinforcement learning for policy optimization at the lower level and multi-task Bayesian optimization for morphology optimization at the upper level. A robustness metric in caging, Minimum Escape Energy, is incorporated into the objectives of both levels to promote caging configurations and enhance manipulation robustness. The evaluation results through four manipulation tasks demonstrate that co-optimizing morphology and policy improves success rates under uncertainties, establishing caging-guided co-optimization as a viable approach for robust manipulation.

Figures (9)

• Figure 1: We introduce CageCoOpt, a framework that jointly optimizes manipulator morphology and control policy for caging-based manipulation. The goal is to push an object (red) to the target region (green). (a) A non-optimized manipulator (blue) struggles with pushing the object reliably. Fingertip poking often fails due to contact slippage, as precise control is required to handle object shape variations and unmodeled dynamics. (b) Through CageCoOpt, the manipulator evolves into a partially "cage"-like morphology, while the policy learns to nudge the object using its inner surface. The manipulator maintains partially caging the object and guides it to the goal without reliance on precise contacts. This co-optimized system adapts to diverse object shapes and external disturbances, ensuring robust manipulation.
• Figure 2: Illustration of energy-bounded caging under gravitational potential energy field $g$. The energy level set is distributed horizontally, such as $\mathcal{L}_{e^{\text{sup}}}(s_{\text{obj}})$. A point-mass object at configuration $s_{\text{obj}} \in \mathbb{R}^2$ lies inside a $\mathbb{R}^2$-bowl (green, $\mathcal{S}_{\text{obs}}$). The escape path $\alpha$ here has the minimum escape energy $\mathcal{E}(s_{\text{obj}})$.
• Figure 3: The proposed CageCoOpt framework.
• Figure 4: Four manipulation tasks detailed in Section \ref{['sec5a']}. Gravity is denoted by $g$.
• Figure 5: Escape paths (green) with minimum escape energy $\mathcal{E}(s_{\text{obj}})$.

Theorems & Definitions (2)

• Definition 1
• Definition 2