Improbable Dexterous Manipulation Research

Hardware

EyeSight Hand

A novel 7-DoF humanoid hand with integrated vision-based tactile sensors for enhanced whole-hand manipulation. Features quasi-direct drive actuation for human-like strength and speed. Evaluated on challenging tasks including bottle opening, plasticine cutting, and plate manipulation. Results demonstrate that tactile feedback dramatically improves task success rates, highlighting the critical role of tactile information in dexterous manipulation.

Project Website

Hardware

DexWrist

A compliant robotic wrist designed to advance manipulation in constrained environments and accelerate policy learning. DexWrist achieves human wrist-like capabilities with mechanical compliance and expanded workspace. It supercharges policy learning through faster teleoperation, shorter task completion, torque transparency for simulation, and enhanced workspace for cluttered scenes.

Project Website

Data Collection

DEXO

A novel hand exoskeleton system designed to teach robots dexterous manipulation in-the-wild through natural and intuitive control. DEXO enables kinematic mirroring and force transparency, allowing human users to directly control a robot's dexterous hand with integrated tactile sensors. The passive exoskeleton design captures high-fidelity interaction data for learning complex tasks, demonstrating significant improvements in task success rates compared to existing teleoperation methods.

Data Collection

DART

A teleoperation platform designed for crowdsourcing robotic data collection through cloud-based simulation and augmented reality (AR). DART addresses scalability limitations of real-world data collection by enabling higher throughput and lower physical fatigue compared to traditional teleoperation. Policies trained on DART-collected datasets successfully transfer to reality and are robust to visual disturbances. All data is automatically stored in DexHub, a cloud-hosted database that will be publicly available for robot learning.

Project Website

Learning and Control

In-hand Reorientation
Demonstration

In-hand Reorientation

Advanced control algorithms for in-hand object reorientation, enabling robots to manipulate objects within their grasp without dropping them. This research focuses on developing robust control strategies that can handle complex object geometries and dynamic reorientation tasks. The system combines tactile feedback, visual perception, and predictive modeling to achieve smooth and precise in-hand manipulation.

Learning and Control

ResiP: Residual for Precise Manipulation

ResiP augments a frozen behavior cloning policy with a closed-loop residual policy trained via reinforcement learning, enabling reliable and precise manipulation. This approach overcomes the limitations of open-loop imitation and direct RL fine-tuning, achieving strong performance on high-precision tasks like alignment and insertion.