Dexterous Manipulation (2026)

Senior scientist, Italian Institute of Technology · Chair of robotics, University of Pisa

Adaptive synergies and compliant, underactuated robotic hands.

The Pisa/IIT SoftHand uses adaptive synergies to obtain many useful grasps with little independent actuation. Its mechanical adaptability highlights a tradeoff: versatile grasping and simple control do not automatically provide independent finger motion for fine manipulation.

Full professor, EPFL

Adaptive coordination and control for grasping under disturbances.

Her work with Farshad Khadivar coordinates fingers through coupled dynamical systems and adaptive torque control. It treats perturbations and uncertain object dynamics as control problems alongside the choice of grasp.

Professor of robotics, TU Berlin

Soft hands that exploit contact and environmental constraints.

RBO Hand research shows how compliant pneumatic fingers can adapt during contact rather than requiring a precise preplanned grasp. This approach makes morphology and the environment part of the control strategy.

Professor, Columbia University · Cofounder, Tangent Robotics

Hand design, tactile manipulation and exploration for dexterous learning.

Sampling-based exploration finds useful states in the constrained hand–object configuration space, then uses them to train dynamic RL policies. This connects classical manipulation geometry to learned control and physical hand experiments.

Professor of mechanical engineering, Stanford University

Grasp taxonomies and the link between hand design and manufacturing tasks.

His grasp-choice study organized manufacturing grasps around task requirements and object constraints. It remains a useful starting point for deciding which hand capabilities a job actually needs.

Professor emeritus, Carnegie Mellon University · Chief scientist, Berkshire Grey

Contact mechanics and manipulation that uses the environment.

Shared Grasping combines precision finger contacts with surfaces in the environment. Its contact-mode planning and hybrid force/velocity control explain how a robot can reposition an object without reproducing every capability of a human hand.

Professor, Carnegie Mellon University · Cofounder and president, FuturHand Robotics

Contact-aware grasp synthesis and retargeting between different hands.

Contact-area retargeting matches useful hand–object contacts across different morphologies. It addresses a central weakness of copying human joint angles: a similar pose can produce a different, or infeasible, grasp on a robot.

Assistant professor, University of Illinois Urbana-Champaign

Optical tactile sensing of contact geometry, deformation and slip.

GelSight measures deformation of an illuminated elastomer through a camera. The resulting contact geometry and deformation signals support force and slip estimation, subject to calibration and sensor mechanics.

Professor, UC Berkeley · Amazon Scholar, Amazon · Advisor, OpenAI

Learning-based dexterity and foundation controllers for robot hands.

DexterityGen uses learned grasp transitions to turn coarse human commands into finer hand control. Its tool-use demonstrations illustrate an assisted low-level controller; they do not establish autonomous execution of every demonstrated task.

Full professor, TU Dresden · Scientific advisor, Kyber Labs

Compact tactile sensors and adaptive in-hand manipulation.

DIGIT made high-resolution optical touch practical in compact robotic fingers. His coauthored motor-adaptation work also shows how recent interaction history can help a hand cope with changing object dynamics.

Principal research scientist, NVIDIA

Large-scale sim-to-real training for dexterous hand and arm control.

DeXtreme transfers agile in-hand reorientation from randomized simulation. DextrAH-G extends the learning stack toward arm–hand grasping with geometric fabrics, a related but distinct task from unrestricted in-hand tool use.

Cofounder and CEO, MyoLab · Adjunct professor, Carnegie Mellon University

Demonstration-augmented RL and dexterous manipulation benchmarks.

Demonstration-augmented policy gradients use expert examples to guide exploration in challenging hand tasks. This line of work helped establish dexterous manipulation as a testbed for combining imitation with reinforcement learning.

Assistant professor, Columbia University

Scalable tactile sensing integrated with learned manipulation.

FlexiTac connects inexpensive flexible tactile hardware to robot learning and simulation. Its piezoresistive sensing stack addresses coverage and integration costs that can limit touch-based dexterity.

Staff research scientist, NVIDIA GEAR

Demonstration generation and bimanual dexterous imitation.

DexMimicGen expands small demonstration sets into simulated bimanual training data. He jointly advised DexMachina, which adds a curriculum for transferring demonstrated object interactions across hands.

Dexterity research lead, Amazon · Affiliate assistant professor, University of Washington

Foundation controllers for dexterity and deployable hand learning.

He coauthored DexterityGen’s grasp-transition controller and assisted manipulation system. Its separation between human commands and learned low-level refinement is useful when evaluating claims about autonomous dexterity.

Associate professor, Carnegie Mellon University · Cofounder and CEO, Skild AI

Accessible learning hands and human-data transfer into robot dexterity.

LEAP Hand made a reproducible learning platform available to more labs. DexWild, led by Tao and Srirama, uses diverse human demonstrations with robot data to improve environmental generalization.

Robotics researcher, Meta Superintelligence Labs · GRAIL lab lead, New York University

Learning from tactile play and accessible dexterous-hand hardware.

T-Dex learns tactile representations from play before combining vision and touch in task policies. RUKA couples an inexpensive underactuated hand with learned transmission maps, showing how hardware and learning can be designed together.

Member of technical staff, Amazon Frontier AI and Robotics

Adaptive in-hand control and dexterous tool manipulation.

Rapid Motor Adaptation uses recent proprioceptive interaction history to adapt in-hand rotation to changing dynamics. DexScrew moves this research toward screwdriving, where changing contacts and tool motion matter beyond a stable grasp.

Associate professor, Stanford University

Human-hand interfaces and functional retargeting for dexterous learning.

DexUMI captures human manipulation through a wearable interface. DexMachina, which she jointly advised, trains robot hands to achieve demonstrated object outcomes rather than relying on geometric pose matching alone.

Assistant professor, Peking University · Founder and CTO, Galbot · Research supervisor, Zhongguancun Academy

Large-scale dexterous grasp datasets and learned grasp generation.

DexGraspNet and its second generation expand training data for multi-finger grasping, including synthetic clutter. Their focus is choosing and executing grasps; this is one component of dexterity rather than evidence of arbitrary tool manipulation.

Research director, Meta Superintelligence Labs · Associate professor, UC San Diego

Tactile and perceptual learning for in-hand manipulation.

Rotating without Seeing learns in-hand object motion from tactile feedback. It makes touch a control signal for occluded contacts, while keeping the evidence scoped to the tested rotation tasks.

Cofounder and head of reinforcement learning, PsiBot

Sequential and bimanual dexterous reinforcement learning.

Sequential Dexterity studies how learned hand skills can be chained into longer tasks. His bimanual benchmark work makes coordination measurable in simulation, with physical deployment remaining a separate question.

PhD researcher, New York University · Research scientist intern, Amazon Frontier AI and Robotics

Tactile play pretraining and learned control for affordable hands.

She led T-Dex and co-led RUKA, combining tactile representation learning with hand hardware. RUKA’s learned transmission maps help control a tendon-driven underactuated mechanism; its teleoperated and autonomous demonstrations should be assessed separately.

Senior research scientist, NVIDIA GEAR

Contact-aware control and human-to-robot manipulation interfaces.

His work spans Shared Grasping’s contact mechanics, DexUMI’s human interface and DexMachina’s functional transfer. This links classical constraints to practical data collection and learning across hand designs.

PhD researcher, Columbia University

Touch-based in-hand learning and scalable tactile hardware.

He co-led Rotating without Seeing and contributes FlexiTac’s tactile hardware/learning stack. These works address both sides of the touch problem: obtaining useful contact signals and learning how to act on them.

Robotics researcher, Honda Research Institute

Model-predictive scaffolding for real-world dexterous RL.

His equal-first-author work uses MPC experience and temporary guidance to initialize and stabilize real-hand learning. The resulting rotation policy runs independently, with externally measured object pose and task-specific setup requirements.

Staff research scientist, Boston Dynamics

Contact-area retargeting of human manipulation across robot hands.

He led contact-area-based kinematic retargeting with Hodgins and Pollard. Its non-isometric correspondence and progressive inverse kinematics preserve useful contacts across different hands; dynamic feasibility and closed-loop robot execution require additional control.

Research engineer, Galbot

Generative dexterous grasping in clutter and from language.

His equal-first contributions to DexGraspNet 2.0 and DexVLG connect local geometry and language instructions to multi-finger grasp generation. This is grasp-selection research, with broader in-hand manipulation still outside those evaluations.

Researcher, University of Illinois Urbana-Champaign · Researcher, Amazon Frontier AI and Robotics

LEAP Hand and data-driven dexterous manipulation platforms.

He led LEAP Hand’s open hardware work and coauthored DexWild. These contributions connect reproducible hand design with experiments on transferring human manipulation data.

AI researcher, Google DeepMind

Human-data co-training for dexterous policy generalization.

His equal-first-author DexWild work combines diverse human data with robot demonstrations. The central contribution is environmental generalization through shared observations and actions, rather than an entirely data-free embodiment transfer.

Incoming PhD researcher, UC Berkeley

Learning dexterity from diverse human interactions in the wild.

As an equal first author of DexWild, he helped bridge human-hand demonstrations and robot policies. The tested transfer distinguishes unseen environments and new arms from the few-shot adaptation needed for a new hand.

PhD researcher, UC Berkeley

Touch-driven rotation and reusable low-level dexterity controllers.

Also published as Alex Z. Yin, he co-led touch-based in-hand rotation and leads DexterityGen. The latter learns reusable grasp transitions while retaining human guidance in its showcased tool tasks.

PhD researcher, Yale University

Tactile-conditioned world models for fine hand-motion prediction.

DTWM conditions a video world model on past touch and hand geometry. Its held-out human-hand forecasting results support improved contact-aware prediction; they do not yet demonstrate a robot policy.

Researcher, Meta Robotics Studio

Functional retargeting for bimanual manipulation of articulated objects.

DexMachina gradually removes virtual object assistance while a policy learns to reproduce human-demonstrated outcomes. Its simulation experiments compare hands with different size and actuation constraints.