Dexterous robotic manipulation is becoming increasingly crucial as robots transition from industrial settings to unstructured human environments, such as household assistance and healthcare support. Achieving adaptability across diverse tasks requires high degree-of-freedom (DoF) robotic hands paired with responsive, autonomous control policies. While non-anthropomorphic robotic hands offer advantages in mechanical simplicity and specialized functionality, their morphological divergence from the human hand presents challenges for intuitive human-in-the-loop control and efficient skill acquisition. This thesis bridges this gap by introducing a modular non-anthropomorphic hand framework and demonstrating how human-driven learning can advance dexterous manipulation.

The foundation of this research is DeltaHands, a modular, high-DoF framework based on the parallel Delta robot mechanism. DeltaHands are designed for high dexterity and ease of fabrication using low-cost, off-the-shelf materials. Their modularity provides a reconfigurable design space, enabling rapid adaptation of the hand’s workspace, material properties, and sensor integration, including in-hand vision and multimodal tactile fingertips. The parallel finger architecture provides an intuitive kinematic structure that simplifies control despite the system’s high dimensionality while benefiting coordinated hand synergies. This framework offers a versatile hand design space for dexterous manipulation.

To address the challenge of controlling non-anthropomorphic hands, we explore human-to-robot motion mapping to leverage human input. Building on the DeltaHands framework, we developed both a vision-based hand-tracking interface and a kinematic-twin interface for direct control. A user study involving non-expert participants across multiple in-hand manipulation tasks demonstrated that the kinematic-twin interface significantly improves success rates and reduces task completion time. Subjective evaluations further indicate that the kinematic twin offers superior usability and lower cognitive workload compared to visual tracking. These results demonstrate that an intermediate device allows users to bypass morphological gaps, effectively unlocking the full dexterity of the robotic hand.

Using high-quality teleoperation demonstrations, we show that a variety of dexterous skills can be acquired through imitation learning. To further reduce data collection effort and improve generalization, we introduce a real-to-sim-to-real learning pipeline that leverages sensorized exoskeleton demonstrations. By capturing direct human–object interactions without a robot in the loop, these demonstrations bootstrap a simulation-based, auto-curriculum reinforcement learning method, which is then transferred to real-world robots in a zero-shot manner. This approach enables the learning of dynamic robotic behaviors using fewer than 15 demonstrations and minimal reward engineering.

Finally, recognizing that tactile sensing is essential for precise manipulation, we design a compact fingertip with integrated multimodal tactile sensors for DeltaHands. We show that integrating static and dynamic contact sensing synergistically enhances manipulation precision, enabling more fine-grained control. We then present an "explore-then-execute" framework that learns task-agnostic object representations from transient tactile signals gathered during exploratory interactions to inform downstream object-centric policies.

In summary, this thesis takes a step towards advancing dexterous manipulation in non-anthropomorphic systems through human-guided adaptation and learning.