GraspADMM: Improving Dexterous Grasp Synthesis via ADMM Optimization

1. Problem Statement

Synthesizing high-quality dexterous grasps for robots with high degrees of freedom (DoF > 20) is a complex challenge requiring the simultaneous satisfaction of four conflicting requirements:

1. Diversity: Broad coverage of object shapes, hand poses, and grasp types.
2. Kinematic Feasibility: Ensuring hand-object contact without physical penetration (collisions).
3. Dynamic Stability: Generating sufficient contact forces to balance gravity and external wrenches (force closure).
4. Speed: Fast synthesis suitable for real-world applications.

The Gap: Existing methods fail to balance these requirements effectively:

• Sampling-based methods (e.g., GraspIt!) ensure diversity but are computationally inefficient for high-DoF hands.
• Gradient-based methods optimize dynamic stability but rely on simplified contact approximations that often result in physical penetrations.
• Simulator-based methods (Dexonomy) achieve good kinematic feasibility and diversity but suffer from a critical flaw: they fix contact points on both the object and the hand during refinement. This rigidity prevents the optimization of dynamic stability metrics (as contact points cannot move to optimal locations) and limits kinematic reachability (if fixed points are unreachable, the grasp fails).

2. Methodology: GraspADMM

The authors propose GraspADMM, a framework that replaces the fixed-point refinement stage of Dexonomy with an Alternating Direction Method of Multipliers (ADMM) optimization pipeline.

Core Innovation: Decoupling Variables

The key insight is to decouple the target contact points on the object ($p_o$) from the actual contact locations on the hand ($p_h$). Instead of forcing $p_h = p_o$ immediately, the algorithm treats them as independent variables and alternates between optimizing them.

The optimization problem is formulated as:
$$\min_{q \in C, p_h, p_o} \max_{\lambda} \mathcal{L}(p_o, p_h, \lambda)$$
Where:

• $q$: Hand pose.
• $C$: The strictly penetration-free kinematic workspace.
• $\mathcal{L}$: The augmented Lagrangian function.
• $e(p_o, n_o)$: The dynamic grasp quality metric (Force Closure).

The ADMM Pipeline (Three Alternating Steps)

1. Update Target Object Points ($p_o$):
   • Goal: Maximize dynamic stability.
   • Method: Perform gradient descent on the grasp quality metric $e$ to move the target points on the object surface to optimal locations.
   • Constraint: Project points back onto the object mesh surface.
2. Update Hand Pose and Points ($q, p_h$):
   • Goal: Achieve kinematic feasibility (reach the targets without collision).
   • Method: Use forward physics simulation (MuJoCo) with transposed Jacobian control. Virtual spring forces pull the hand contact points ($p_h$) toward the shifted targets ($p_o - \lambda$).
   • Advantage: The physics simulator naturally resolves exact collisions, ensuring $q \in C$ (no penetration) without needing differentiable collision physics.
3. Update Dual Variable ($\lambda$):
   • Goal: Enforce consistency between $p_h$ and $p_o$.
   • Method: Standard ADMM dual ascent ($\lambda \leftarrow \lambda + p_h - p_o$).
   • Stability Mechanism: A dual-variable annealing strategy resets $\lambda$ to zero if it exceeds a threshold, preventing divergence caused by the non-linear constraints.

3. Key Contributions

1. Novel ADMM Formulation: First application of ADMM to dexterous grasp synthesis, successfully decoupling dynamic metric optimization from hard kinematic collision constraints.
2. Contact Mobility: Unlike previous simulator-based methods, GraspADMM allows contact points to move locally on both the object and the hand, actively optimizing for stability while respecting physical boundaries.
3. Robustness to Low Friction: The method explicitly optimizes for force closure, making it significantly more robust in extreme low-friction environments ($\mu = 0.1$) where other methods fail.
4. Zero Penetration: By leveraging a physics simulator for the kinematic step, the method guarantees zero penetration depth, a common failure point in gradient-based approaches.

4. Experimental Results

The authors evaluated GraspADMM on a benchmark of 5,697 objects using the Allegro and Shadow hands, comparing against state-of-the-art baselines (DexGraspNet, FRoGGeR, SpringGrasp, BODex, and Dexonomy).

• Type-Unaware Synthesis:
  • Achieved a 74.6% Grasp Success Rate (GSR), a ~14% absolute improvement over the previous best (Dexonomy at 60.5%).
  • Maintained 0.0 mm penetration depth and improved diversity metrics.
• Type-Aware Synthesis:
  • Demonstrated a ~100% relative improvement in GSR across Power, Intermediate, and Precision grasp types compared to Dexonomy (e.g., Power grasp GSR increased from 32.9% to 62.8%).
  • Achieved higher Contact Link Numbers (CLN), indicating richer multi-finger contacts.
• Low-Friction Robustness:
  • Under extreme low-friction conditions ($\mu = 0.1$), Dexonomy struggled to find even a single stable grasp out of 100 attempts for many objects, whereas GraspADMM consistently generated diverse, stable grasps.
• Efficiency:
  • Processing speed is approximately 19 samples/second (slower than Dexonomy's 24 s/s due to iterative ADMM, but faster than pure gradient methods).
  • The higher success rate results in a comparable or higher yield of usable grasps per second.

5. Significance

GraspADMM represents a significant paradigm shift in dexterous manipulation planning. It bridges the gap between sampling-based diversity, simulator-based kinematic safety, and gradient-based dynamic optimization. By solving the "fixed contact" limitation of previous simulator-based methods, it enables robots to generate grasps that are not only physically possible (no collisions) but also dynamically robust (stable under gravity and friction). This is crucial for deploying dexterous hands in unstructured, real-world environments where object friction and geometry vary widely.