UNO Push: Unified Nonprehensile Object Pushing via Non-Parametric Estimation and Model Predictive Control

TL;DR

The paper tackles precise pushing under perception and physics uncertainties by proposing UNO Push, a data-efficient framework that unifies non-parametric estimation of the planar push transition with closed-loop MPC. It demonstrates that accurate object manipulation can be achieved from a small number of exploratory actions, with online adaptation enabling transfer across novel objects. Key contributions include a Gaussian Process-based estimation of the push transition and its inverse, online updates to refine predictions, and MPC-driven action generation that yields millimeter-level precision without heavy object-specific training. Empirical results on a real 7-DoF robot show performance comparable to baselines trained with thousands of samples, validating cross-object transfer and robustness to perturbations. Overall, UNO Push offers a practical, generalizable low-barrier solution for nonprehensile manipulation tasks in real-world settings.

Abstract

Nonprehensile manipulation through precise pushing is an essential skill that has been commonly challenged by perception and physical uncertainties, such as those associated with contacts, object geometries, and physical properties. For this, we propose a unified framework that jointly addresses system modeling, action generation, and control. While most existing approaches either heavily rely on a priori system information for analytic modeling, or leverage a large dataset to learn dynamic models, our framework approximates a system transition function via non-parametric learning only using a small number of exploratory actions (ca. 10). The approximated function is then integrated with model predictive control to provide precise pushing manipulation. Furthermore, we show that the approximated system transition functions can be robustly transferred across novel objects while being online updated to continuously improve the manipulation accuracy. Through extensive experiments on a real robot platform with a set of novel objects and comparing against a state-of-the-art baseline, we show that the proposed unified framework is a light-weight and highly effective approach to enable precise pushing manipulation all by itself. Our evaluation results illustrate that the system can robustly ensure millimeter-level precision and can straightforwardly work on any novel object.

Figures (17)

• Figure 1: A robot manipulator is tasked to manipulate an unknown object to trace reference trajectories (blue). Without analytically modeling the contacts or other physical components of the system, our UNO Push framework enables precise manipulation (yellow trajectories) by pushing the object with the gripper of the robot (top), or by an unknown object grasped by the gripper (bottom).
• Figure 2: Left: The representation of the control, through two angles $\alpha_t$ and $\beta_t$ in the object's body frame; Middle: The rigid body motion of the object $\leftidx{^b}{g}_t$; Right: The smoothened execution strategy of control. Instead of retreating the gripper back to the point $P$, the robot moves the gripper to a point $P'$ closer to the object.
• Figure 3: Three example data points collected by pushing a cylinder through random controls. The red dashed and solid arrows represent the x-axis of the object's body frame before and after the push, respectively. The object's configuration has been changed through translation $\Delta \mathbf{x}$ and rotation $\Delta\phi$. In our experiments (Sec. \ref{['sec:experiments']}), we applied the model learned on the cylinder object to directly manipulate other objects.
• Figure 4: Trajectory simulation and optimization by MPC. By iteratively propagating the object's configuration with the estimated models $\Gamma$ and $\Gamma^{-1}$ and random perturbations, a bunch of trajectories (blue) are simulated to a horizon $L$. The optimal one (thick blue), which is closest to the reference trajectory (dashed), is selected to extract the control input for execution.
• Figure 5: The objects used in the experiments. A. Bowl (YCB $\#0024$) B. Wood block (YCB $\#0071$) C. Windex bottle (YCB $\#0022$) D. Sugar box (YCB $\#0004$) X. A Cylinder object