Abstract
Manipulating deformable objects, such as dough, presents a significant challenge in robotics due to the high-dimensional state space and complex dynamics involved. This paper introduces DGform, a novel framework that utilizes dynamic heterogeneous graphs to represent the interaction between a bimanual robot and a deformable object. By abstracting the dough and the robot's end-effectors into a unified graph structure, the system learns goal-oriented manipulation policies from human demonstrations and simulation. Experimental results demonstrate that DGform significantly outperforms existing reinforcement learning baselines in both simulation and real-world deployment on a humanoid robot.
Introduction
The manipulation of soft, deformable objects—often referred to as a "grand challenge" in robotics—differs fundamentally from rigid body manipulation. Dough, for instance, possesses infinite degrees of freedom; it stretches, squishes, and deforms unpredictably under pressure. Traditional robotic approaches struggle with two primary hurdles: state representation and policy learning. Raw pixel data is often too noisy and high-dimensional for effective control, while particle-based tracking is computationally prohibitive for real-time applications. Furthermore, Reinforcement Learning (RL) from scratch is time-consuming, and Learning from Demonstration (LfD) often lacks the robustness to generalize beyond the specific trajectories shown by a human.
To address these challenges, we present DGform, a framework designed to bridge the gap between simulation and reality through a compact, graph-based representation that captures the essential dynamics of bimanual dough rolling.
The DGform Framework
Dynamic Heterogeneous Graph Abstraction
At the heart of DGform is a specialized graph representation. Rather than attempting to track every particle of the dough, the system abstracts the object into a 9-node graph consisting of one central point and eight boundary points. This sparse representation captures sufficient geometric and semantic information for the rolling task while remaining computationally efficient.
Crucially, the graph is heterogeneous. It incorporates "manipulator nodes" representing the robot's hands alongside the dough nodes. This allows a Graph Convolutional Network (GCN) to explicitly reason about the spatial relationships and contact dynamics between the robot and the object. As the dough changes shape, the graph edges update dynamically, reflecting the evolving state of the system.
Goal-Oriented Policy Learning
The policy is conditioned on a "goal graph," which represents the desired final shape of the dough. The robot's objective is to minimize the discrepancy between the current graph state and the target state. To accelerate the learning process, the framework utilizes Lagrange-based Imitation Guidance. This method incorporates human demonstrations but avoids rigid mimicry. By using a self-adjusting Lagrange coefficient, the system treats human data as a flexible guide: it provides strong direction during early training phases but allows the robot to explore and optimize its own trajectories as it gains proficiency.
Bimanual Coordination and Control
To ensure fluid, collision-free movement between the two arms, DGform employs Bimanual Relative Parameterization (Bi-RP). This approach coordinates the high-level waypoints generated by the policy. These waypoints are then processed through Linear Quadratic Tracking (LQT) to produce smooth, 1000Hz motion commands suitable for real-world hardware.
Experimental Setup
Simulation and Data Collection
The model was initially trained in PlasticineLab, a differentiable physics simulator tailored for soft body dynamics. To ground the model in human-like movement, the team collected a multimodal dataset of human dough rolling. This involved an Optitrack motion capture system to record hand trajectories and ZED 2i RGB-D cameras to capture the resulting dough deformations.
Real-World Deployment
The trained policy was transferred to the CURI robot, a humanoid platform equipped with two 7-DoF arms. Because the graph-based representation is consistent across both simulation and reality (extracted from vision data in both cases), the "reality gap" was significantly minimized, allowing for successful sim-to-real transfer.
Results and Discussion
DGform was evaluated against several baselines, including standard Proximal Policy Optimization (PPO) and DiffSkill. The results highlight the advantages of the graph-based approach:
- Performance Metrics: DGform achieved a reward score of 756.1, surpassing the best PPO baseline of 629.5.
- Shape Accuracy: In terms of Intersection over Union (IoU)—measuring the overlap between the final dough shape and the goal—DGform reached 0.5573, the highest among all tested methods.
- Generalization: The model demonstrated high robustness. It maintained stable performance even when subjected to variations in camera positioning or initial dough shapes, conditions that typically cause pixel-based methods to fail.
Conclusion and Future Work
DGform demonstrates that dynamic heterogeneous graphs provide a powerful abstraction for complex bimanual tasks involving deformable objects. By focusing on semantic nodes rather than raw pixels, the framework achieves efficient learning and robust sim-to-real transfer.
Future iterations of this work will aim to expand the model into three dimensions. By incorporating 3D graph neural networks and point cloud data, the researchers hope to address more complex deformations and mitigate issues related to occlusion, such as when a rolling pin hides the dough from the camera's view. This work represents a significant step toward robots capable of performing intricate culinary and industrial tasks.
Summary updated