Capture the skill.
Teach the feel.
Start with data. Validate on one task. Extend to more robots. Two connected disciplines turn human contact intelligence into something machines can learn: acquisition and learning.
The problem begins at contact.
When a connector catches, a skilled worker reads the force and adjusts the angle. When a hose pulls tight, they change the supporting hand. That judgment lives in what they feel and how they respond, as much as in the path they follow.
We develop both the interfaces that let people demonstrate their natural skill and the frameworks that learn physical interaction from those records. From hardware that stays out of the way to representations that preserve what force means.
Acquisition / The skill in arms and wrists
Kinesthetic skill acquisition. Recording the ability to work through force and movement requires three things to come together.
- 01
A stable foundation for teaching
We manage energy in the bilateral loop so demonstrations can continue through communication delays and hard contact. Time-domain passivity research guides our validation of stability conditions and real hardware behavior.
- TDPA
- Bilateral control
- Contact transitions
- 02
Hardware that lets the hand lead
Low inertia, low friction, and compensation control aim to reduce the device’s own resistance. We design the bimanual workspace and handle placement together so operators spend less effort working against the hardware.
- Hardware development
- Bimanual coordination
- Friction compensation
- 03
Feel the surface. Move freely.
Hard surfaces should feel distinct. Free motion should feel natural. We study stiffness rendering, delay, and force distortion together to help operators distinguish subtle changes in contact.
- High-stiffness rendering
- Transparency
- Energy dissipation
Research behind acquisition
Explore the thinking and methods behind the research.
- Data acquisitionPublic overview / Details by invitation
Stable contact teaching: managing energy in bilateral control
Building on Prof. Jee-Hwan Ryu’s time-domain passivity research, we explore what keeps demonstrations stable through hard contact and communication delays.
- Data acquisitionPublic overview / Details by invitation
Firm surfaces. Natural movement.
Virtual inertia and UVI research for high-stiffness rendering and transparency, and methods for evaluating realism.
- Data acquisitionPublic overview / Details by invitation
Bimanual haptics that let human hands lead
Design criteria that bring together bimanual workspace, force capability, inertia, and friction.
Acquisition / Touch, without getting in its way
Tactile skill acquisition. Contact location, pressure, and slip all change how we work. Our sensing glove aims to capture that information while preserving the operator’s feel for texture and stiffness.
Our development goal is greater task realism than existing published gloves. We consider thickness, flexibility, surface friction, and comfort together, and plan comparisons against bare-hand task time, success, and force use. Comparative performance has not yet been established.
People should demonstrate their own skill, without adapting to an awkward glove. That is why we evaluate the naturalness of the demonstration alongside sensor resolution.
Representation / Contact and response, kept together
We connect task video, robot commands and actual motion, operator force and stiffness adjustment, environmental contact force, and fingertip touch. Attention to objects and contact points adds context.
We preserve each sensor’s original timing and distinguish measurements from estimates and annotations. Records include successful motion as well as retreats, retries, and the task stages they belong to.
Learning / Representations that understand force
Beyond turning tactile input into images for a vision encoder, we study frameworks that learn the relationship between force history and action.
Learning from haptic data
Contact direction, magnitude, change, slip, and operator force adjustment become physically meaningful representations. The goal is to learn when force was applied, why, and how it changed the next action.
- Physical representations
- Temporal resolution
- Contact state
Connecting VLAs to contact response
We pair contact observations and responses with VLAs that understand objects and tasks. Using the same task and data budgets as current baselines, we evaluate success, damage, recovery, and completion time.
- VLA
- Policy learning
- Field evaluation
World models informed by human contact strategies
We explore predictions of contact state, force change, and action outcomes alongside visual futures. The research direction combines force data from physics engines with records of how real operators respond to resistance and uncertainty.
- World models
- Predicting action outcomes
- Human contact strategies
Research into learning contact
Explore the thinking and methods behind the research.
- LearningPublic overview / Details by invitation
From force history to action: learning with haptic data
Research directions for policies that understand force and touch, and world models grounded in contact.
- LearningPublic overview / Details by invitation
Different signals. One decision.
How signals on different clocks become one contact-aware decision.
One task today. Broader intelligence ahead.
We start with tasks we can define and evaluate, such as connector assembly. We vary parts and conditions, study failures, and collect the data needed to improve. Task-specific evaluation criteria connect through Dream Set to records for learning.
We aim to bring intelligence to existing robots and equipment. Integration requires review of control interfaces, sensors, and safety needs. Haptic representations and world models are development directions, not claims of a finished, validated general-purpose product.
From research to real tasks.
See how the research connects to acquisition hardware, datasets, and field applications.