DITTO-X makes a robot hand align the operator before takeover
DITTO-X runs an actuated hand exoskeleton in both directions, rendering robot contact to a human and matching the operator to the robot before intervention.

Stanford and Columbia researchers released DITTO-X on September 30, 2026. The actuated exoskeleton sends a person's finger motion to three commercial dexterous hands, renders hand-dependent force and contact feedback, and can reverse direction so the operator follows the robot's commanded finger configuration before control transfers. In a six-person study, that matched takeover improved a contact-rich tool-use intervention from 27.7% to 78.3% success against a tracking-glove baseline. Paper and version record.
Explore tactile datasets for the broader comparison.
Key takeaways
- Forward teleoperation renders hand-dependent force feedback. The evaluated Sharpa configuration also supplies fingertip vibration; that contact-event channel is not listed for Wuji 2 or Inspire.
- Reverse teleoperation maps the robot's commanded hand configuration back to the exoskeleton before takeover, reducing the pose jump that can release a held object.
- The project describes 1,888 episodes and more than 16 hours at 30 Hz, but its Code, Dataset and Hardware Guide controls all still say “coming soon.”
How bidirectional control works
DITTO-X aligns actuator axes with human finger joints. Index and middle fingers can map joint-to-joint on the 22-DoF Sharpa and 20-DoF Wuji 2 hands. Thumb motion is retargeted in task space, and the six-DoF Inspire hand receives a lower-dimensional flexion mapping. The exoskeleton independently tracks thumb, index and middle fingers; robot ring and little fingers are coupled to the middle-finger motion. Feedback differs by hand: Table I lists contact-event vibration for Sharpa only. Motor-current-derived torque is an estimate, not a measured surface pressure map. Hardware and mapping details.
During autonomous execution, the robot's current command is projected back onto the exoskeleton. A keyboard request initiates the takeover process; actual transfer is gated by departure of the human joints from the independently maintained projection. Following the command does not prove that the physical robot achieved that pose. Safety measures reported by the authors include software joint limits, motor-current limits below 200 mA with torque no greater than 0.08 N·m, and an emergency stop.
Results under the reported conditions
Six participants completed blinded size and compliance discrimination, regular tool-use teleoperation and mid-policy intervention. With both feedback modes, size discrimination is 86.1% and compliance discrimination 91.7%, against 33.3% chance. Removing either force or vibration reduces the scores; the experiment supports complementarity within this prototype, not a universal ranking of haptic modalities.
For collecting tong-use demonstrations, DITTO-X succeeds on 70.0% of trials versus 46.7% for the Manus Pro tracking glove, a 23.3-point gap. Mean time per success falls from 85.5 to 56.1 seconds. In intervention, success is 78.3% versus 27.7%, while time per success falls from 97.4 to 35.7 seconds. The size/compliance tasks use six trials per condition per participant; tool-use and intervention protocols specify ten. These are repeated measurements from six people. The project reports intervention success as a mean of participant rates, not a pooled numerator out of 60. Intervention begins with the tool already grasped, whereas full collection includes acquiring it.
Policy tests use 30 paired starting configurations per task. Before DAgger, DITTO-X data produces final-stage success of 63.3% on tong use, 63.3% on raspberry placement and 40.0% on battery insertion; Manus data yields 13.3%, 43.3% and 36.7%. After two DAgger rounds retaining 20 successful interventions per round, the DITTO-X policies reach 86.7%, 93.3% and 90.0%. The paper also quantity-matches extra demonstrations. Failed interventions are excluded from training; the policies encounter interface-specific failure states. This comparison addresses data volume but does not isolate every cause of the interface advantage.
RoboSkin analysis
The most important change is at the authority boundary. Conventional human intervention switches the controller, but the operator may begin with fingers in a different pose from the robot. With a dexterous hand already holding a tool or fragile object, that mismatch is itself a disturbance. DITTO-X treats body alignment as part of robot teleoperation, while force and contact feedback give the operator evidence that cameras can lose under occlusion.
The interface also shows why haptics should be evaluated downstream. Better blind discrimination is useful, but the stronger evidence is that contact-aware demonstrations produce better autonomous policies and that matched interventions cover failure states. This connects human feedback to the broader robot learning pipeline.
Limitations and availability
DITTO-X is an arXiv v1 preprint and RoboSkin.ai has not reproduced its hardware or results. The main user study has six participants, most policy experiments use the Sharpa hand, and mappings still require known hand kinematics and per-hand scaling. The design is not demonstrated on full-body teleoperation or unknown hands.
The official page says it releases the DITTO-Human Dataset: 847 teleoperated, 379 human-intervened and 662 autonomous episodes, totaling 1,888. The October 2 check found Code, Dataset and Hardware Guide controls marked coming soon. The November 4 project HTML and scripts still expose those controls without working artifact links. No downloadable episodes, repository or license was verified. The paper's CC BY-NC-ND 4.0 license does not license absent code or hardware files.
What must an intervention log preserve?
A takeover request and an executed takeover are separate events. For a pilot on another hand, preserve the policy command, measured robot joints, projected human configuration, measured operator joints, request time and actual authority change. Record which feedback channels the operator received. Otherwise a smooth video cannot show whether an alignment delay, tracking error or missing cue caused a failure.
The study protocol also separates the six-person interface experiment from policy data collected by one proficient operator. Policy evaluations use 30 paired initial states per task, including at least ten out-of-distribution placements. Wuji and Inspire demonstrate additional hand mappings, but their DAgger evaluations remain future work. Prototype current limits and an emergency stop do not constitute a safety certification.
For the next collection campaign, retain attempted, accepted and discarded corrections with reasons. Evaluate a frozen policy on independently specified starts, and report operator assistance separately. This is a proposed collection record, not a claim that RoboSkin has validated DITTO-X hardware. The teleoperation signal-role comparison explains why operator haptics and policy tactile observations need different fields.
Sources and related resources
- DITTO-X v1, submitted September 30, 2026
- Full user study, policy evaluation and safety details
- Official project and dataset composition
- RoboSkin tactile feedback guide
*Evidence review updated November 4, 2026. Original publication dates are retained. We inspected sources and public metadata; we did not run these robot systems or download their training archives.*


