DITTO: Dexterous Interface for Transparent TeleOperation

Overview

Abstract

Collecting data for manipulation with high-DOF hands is challenging, as interfaces must capture rich hand motion while rendering the contact interactions essential for precise manipulation. Existing data collection approaches face a trade-off: teleoperation ensures deployment consistency but lacks force feedback, while handheld (in-the-wild) systems provide natural force transparency but introduce a visual embodiment gap at deployment. We present DITTO, a Dexterous Interface for Transparent TeleOperation, which resolves this through the anatomically informed co-design of a dexterous 7-DOF robotic hand and a kinematically equivalent motorized exoskeleton. A 1-to-1 actuator mapping between the exoskeleton and robotic hand enables handheld (in-the-wild) data collection and bilateral teleoperation with joint-level force feedback unified in a single platform. We demonstrate that the DITTO exoskeleton spans the operator's natural index-to-thumb workspace, and showcase DITTO's dexterous capabilities via learned policies on contact-rich tasks.

Force Feedback

DITTO's leader-follower 1:1 joint correspondence enables joint-level force feedback.

Dexterous Teleoperation

DITTO enables highly dexterous teleoperation, leveraging coordinated index/thumb control.

x1

Fanning a deck of cards to pick up only the top card.

x1

Precise alignment to insert a nut into a bolt, followed by fast threading through dynamic motion.

x1

Grasping two objects with one hand at the same time.

x1

Opening an AirPods case and extracting one earbud.

Handheld (In-The-Wild) Data Collection

DITTO supports both:

Bilateral teleoperation with force feedback, for high-fidelity, robot-aligned data.
Handheld collection with the exoskeleton alone, for intuitive and scalable data collection.

x1

Autonomous Policies

Policies trained on data collected via DITTO teleoperation perform contact-rich tasks relying on the hand's intrinsic degrees of freedom.

x1

Turns a key until the lock engages, applying torque with the fingers alone and minimal wrist motion.
Sound on

x1

Reorients a mahjong tile through two consecutive flips between the fingers, then places it in a target box.

Picks up a retractable USB drive, slides the plug out while keeping its grasp, and inserts it into a slot.
*The full-frame view is the policy's side camera, which is also a policy input.

x1

Bimanual: picks up a Rubik's Cube and rotates its right and top faces with thumb-index manipulation.

Success rate over 20 trials per stage, for Diffusion Policies (Transformer) trained on DITTO teleoperation data.

User Study

In a user study with 8 participants, DITTO's human-aligned kinematics and force feedback led to faster manipulation and gentler contact forces.

Experiment 1: Mahjong Tile Flipping

Flip a mahjong tile as many times as possible in 60 s, testing speed during coordinated index–thumb motion.

Two frames of the DITTO hand flipping a mahjong tile on a platform mounted on a force-torque sensor; arrows mark the flip direction.

With the DITTO exoskeleton and force feedback, participants flipped the tile 2.2 times as often as with fingertip retargeting, and 1.3 times as often as with feedback turned off.

Stacked bars of tile flips per participant (S1 to S8) and their average, for DITTO, DITTO-NF and Fingertip, with one shade per trial. — Flips per participant (S1–S8), summed over three trials; shading marks the trial. *Avg* is the mean ± std across the 8 participants (paper Fig. 3b).

Every participant flipped the most tiles with the DITTO exoskeleton, and all but one did best with force feedback on. DITTO beats Fingertip (p ≤ 0.023, one-sided permutation test). Flips are each participant's total over the three trials; bars show the mean ± std across the 8 participants.

Experiment 2: Ratchet Turning

Turn a ratchet three full revolutions as fast as possible (instrumented with a force-torque sensor).

The DITTO hand gripping a ratchet mounted on a force-torque sensor; arrows mark the turning direction.

The robot arm is locked in place for this task, so any force on the sensor comes from the fingers alone.

Users took 30% less time using DITTO compared to fingertip retargeting. For this metric, performance was similar with and without force feedback.

With DITTO force feedback, users applied 47% less force, without significantly impacting task throughput.

This shows DITTO prevents undesired excessive forces that do not lead to better task performance but could damage the hardware.

With DITTO, 92.8% of force samples stayed under 14.8N (which is Fingertip's average contact force).

One-sided permutation tests: DITTO vs. DITTO-NF on both force metrics, p = 0.012; DITTO vs. Fingertip on every metric, p ≤ 0.023.

Empirical CDF of ratchet contact force for DITTO, DITTO-NF and Fingertip; the DITTO curve rises fastest, reaching 1 by about 28 N. — Share of force samples below each force level. Bold lines are the average across all 8 participants, thin lines are individual participants; further left means gentler contact (paper Fig. 3a).