Primitive-Informed Sampling-Based MPC for Multi-Fingered Dexterous Manipulation

The search context does not contain information about a framework explicitly titled "Primitive-Informed Sampling-Based MPC." However, it details related approaches that achieve efficient multi-fingered dexterous manipulation by integrating low-dimensional manipulation primitives or synergies with control methods:

• PCA + ProMP Framework: A study published in Frontiers in Robotics and AI (2026) applies Principal Component Analysis (PCA) and Probabilistic Movement Primitives (ProMPs) to capture low-dimensional finger-joint synergies. This approach is combined with existing motion planners to regulate local contact, demonstrating robust execution on both simulated and physical humanoid hands with an ~82% success rate.
• Latent Space Sampling: Research on biomimetic hands integrates VQ-ACE (Vector Quantized Autoencoder with Continuous Embedding) with Model Predictive Control (MPC). This method performs latent space sampling to reduce action space dimensionality, producing more human-like behavior and higher success rates in tasks like ball rolling and object picking compared to standard MPC.
• Sampling-Based MPC with VLMs: Separate work demonstrates sampling-based MPC on tendon-driven hands using MuJoCo and Visual Language Models (VLMs) for objective function adaptation, achieving tasks like ball flipping and catching without extensive training, though it does not explicitly rely on predefined low-dimensional primitives.

While the specific "Primitive-Informed" title is not present, the described methodology aligns with the PCA + ProMP synergy-based representation and latent space sampling techniques found in the provided context.

Overview. The material addresses a core challenge in multi-fingered dexterous manipulation: planning and control in high-dimensional, contact-rich joint spaces where sampling-based model-predictive control (MPC) can become computationally expensive and prone to ineffective exploration. It proposes a primitive-informed sampling-based MPC framework in which candidate control actions are not generated uniformly over the full manipulation space, but are instead biased by low-dimensional manipulation primitives. These primitives encode structured, task-relevant behaviors—such as coordinated finger motions or grasp-like action patterns—that can be sampled or parameterized more efficiently than raw joint commands.

Key contributions. The central contribution is a planning architecture that combines the flexibility of sampling-based MPC with the efficiency of manipulation primitives. By using primitives as a prior over useful action sequences, the method reduces the effective dimensionality of the search space while still allowing MPC to evaluate and correct candidates against a dynamics model or cost function. This is particularly valuable for dexterous hands, where naive sampling over many joints and contact constraints may produce infeasible, low-quality, or slow-to-converge plans. The approach effectively turns primitives into a structured exploration mechanism, enabling the controller to focus on plausible multi-fingered actions rather than exploring the full action space indiscriminately.

Significance. The work matters because dexterous manipulation remains difficult to control in real time due to high degrees of freedom, nonlinear dynamics, and dense contact interactions. Primitive-informed MPC offers a practical middle ground between fully learned policies, which can be brittle or opaque, and generic sampling-based planners, which can be inefficient in high-dimensional spaces. By leveraging low-dimensional behavioral priors, the framework improves the likelihood of finding useful control sequences quickly, making it a promising direction for scalable, real-time control of advanced robotic hands.