ManiDP: Manipulability-Aware Diffusion Policy for Posture-Dependent Bimanual Manipulation

TL;DR

ManiDP addresses posture-dependent demands in postural bimanual manipulation by learning bimanual manipulability ellipsoids and encoding them on SPD manifolds, then guiding a diffusion-based policy with these posture priors. It introduces BAM for symmetric and BRM for asymmetric coordination, uses SPD-GMM/SPD-GMR to learn BMEs, and performs manipulability-guided diffusion sampling to produce task-compatible dual-arm trajectories. Across six real-world tasks, ManiDP shows substantial improvements in manipulation success rate and task compatibility compared to baselines, demonstrating the practical value of posture-aware priors in bimanual diffusion. This work advances dexterous manipulation by integrating high-level task cognition about posture into generative policy learning.

Abstract

Recent work has demonstrated the potential of diffusion models in robot bimanual skill learning. However, existing methods ignore the learning of posture-dependent task features, which are crucial for adapting dual-arm configurations to meet specific force and velocity requirements in dexterous bimanual manipulation. To address this limitation, we propose Manipulability-Aware Diffusion Policy (ManiDP), a novel imitation learning method that not only generates plausible bimanual trajectories, but also optimizes dual-arm configurations to better satisfy posture-dependent task requirements. ManiDP achieves this by extracting bimanual manipulability from expert demonstrations and encoding the encapsulated posture features using Riemannian-based probabilistic models. These encoded posture features are then incorporated into a conditional diffusion process to guide the generation of task-compatible bimanual motion sequences. We evaluate ManiDP on six real-world bimanual tasks, where the experimental results demonstrate a 39.33$\%$ increase in average manipulation success rate and a 0.45 improvement in task compatibility compared to baseline methods. This work highlights the importance of integrating posture-relevant robotic priors into bimanual skill diffusion to enable human-like adaptability and dexterity.

Figures (6)

• Figure 1: Illustration of posture-dependent bimanual task requirements. (a) Humans naturally adjust their arm posture to optimize manipulation performance. Manipulability serves as a key metric to represent this task-related posture adaptation. (b) Robots should therefore have manipulability awareness to satisfy posture-dependent task requirements.
• Figure 2: Overview of the Manipulability-Aware Diffusion Policy (ManiDP). (a) Unlike the typical diffusion policy that focuses solely on trajectory imitation, ManiDP incorporates BMEs learning into the skill diffusion process to refine dual-arm configuration, enabling enhanced compatibility with posture-dependent task requirements. (b) ManiDP learns a joint distribution of BMEs features and robot trajectories through two key steps: 1) bimanual manipulability learning, where expert BMEs are learned and reproduced using GMM and GMR on the SPD manifold, and 2) manipulability-guided diffusion sampling, where ManiDP iteratively refines the generated bimanual trajectories by minimizing the SPD objective between the current BMEs and the reproduced expert BMEs.
• Figure 3: Visualization of BMEs on the SPD manifold $\mathcal{S}_{++}^2$. BMEs follow a geodesic path (green curve) on the cone-shaped $\mathcal{S}_{++}^2$, contrasting with the Euclidean path (red dashed line). Each geodesic point represents a bimanual manipulability matrix with an associated tangent space. SPD-GMM and SPD-GMR learn BMEs using three Riemannian operations: logarithmic map $\text{Log}(*)$, exponential map $\text{Exp}(*)$ and parallel transport $\Gamma(*)$.
• Figure 4: BMEs learning results of Tower Hanging task on SPD manifold. ManiDP can generate geometrically plausible BME sequences within the cone-shaped SPD manifold by considering the intrinsic Riemannian properties of BMEs.
• Figure 5: BMEs learning results of Tower Hanging task on the time domain. The results show that ManiDP can capture the key states of expert BMEs and then accurately reproduce them at each time step.