# Can One System Unify Handheld Data Collection and Bilateral Teleoperation for Dexterous Hands?

Columbia University researchers say yes. A paper published today on arXiv introduces DITTO — Dexterous Interface for Transparent TeleOperation — a system that resolves a fundamental tension in dexterous manipulation data collection by co-designing a **7-DOF robotic hand** and a kinematically equivalent motorized exoskeleton around human anatomy. The key engineering decision: a 1-to-1 actuator mapping between exoskeleton and robotic hand, which allows the same platform to serve as both a handheld, in-the-wild capture device *and* a bilateral teleoperation interface with joint-level force feedback. The team — Joaquin Palacios, Katelyn Lee, Cheng Zhang, Zhanpeng He, and Matei Ciocarlie — demonstrate that the DITTO exoskeleton spans the operator's natural index-to-thumb workspace, and validate learned policies on contact-rich tasks. For the humanoid industry, where [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) remains the hardest unsolved problem in deployment, this architecture directly attacks the data bottleneck that every hardware team — from [Figure AI](https://humanoidintel.ai/companies/figure-ai) to [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) — is trying to crack.

---

## The Core Problem: Why Dexterous Hand Data Is So Hard to Collect

The manipulation data collection problem for high-[degree-of-freedom](https://humanoidintel.ai/glossary/degrees-of-freedom) hands splits into two unsatisfactory camps, and the DITTO paper articulates this tradeoff more cleanly than most prior work.

**Camp one: traditional teleoperation.** The operator controls the robot remotely, so deployment consistency is high — the robot sees what it will see during autonomous execution. The fatal flaw is the absence of force feedback. When you're manipulating a compliant object, threading a cable, or picking up a fragile item, you rely on tactile and proprioceptive signals that standard teleoperation simply does not convey back to the operator. The resulting demonstrations are impoverished — the operator either over-grips or drops objects, and the collected trajectories fail to encode the subtle force modulation that contact-rich tasks demand.

**Camp two: handheld (in-the-wild) systems.** Here the operator directly handles objects with their own hand, often instrumented with gloves or motion capture markers. Force transparency is natural — you feel everything. But the embodiment gap at deployment is severe. The robot hand has a different kinematic structure, different contact geometry, and different visual appearance than a human hand. Policies trained on this data must bridge a substantial visual and proprioceptive domain gap when transferred to the actual hardware.

DITTO's proposed resolution is anatomically informed co-design: build the robotic hand and the exoskeleton together, from the start, so they are kinematically equivalent. The 1-to-1 actuator mapping means every joint state in the exoskeleton corresponds directly to a joint state in the robot hand, with no kinematic remapping layer that could introduce latency or approximation errors.

---

## What "Kinematically Equivalent" Actually Means Here

The phrase deserves unpacking because it carries significant engineering weight. A kinematically equivalent exoskeleton is not merely a device that roughly tracks hand pose — it is one where each actuated joint in the worn device corresponds bijectively to an actuated joint in the robot. This has two immediate consequences:

1. **Bilateral force feedback becomes tractable.** Because the mapping is 1-to-1, force measured at any robot joint can be reflected back to the corresponding exoskeleton joint without solving an ill-posed Jacobian inversion problem. The operator feels, at the joint level, what the robot feels — which is the closest you can get to in-the-hand force transparency while still using robot hardware.

2. **Handheld data collection closes the embodiment gap.** When the operator holds objects while wearing the exoskeleton, the robot hand (physically attached or virtually mirrored) shares the same kinematic configuration. The visual appearance at deployment matches the data-collection setup because the data *was* collected with the robot hand, not a human hand proxy.

The paper specifies the system as a **7-DOF robotic hand** with a motorized exoskeleton that spans the operator's natural index-to-thumb workspace. The authors validate dexterous capabilities through learned policies on contact-rich tasks — the canonical benchmark for whether a manipulation system is actually useful.

---

## Why This Matters for Humanoid Deployment

Every major humanoid program is currently bottlenecked on manipulation data quality, not locomotion. Walking is largely solved at the research level; [whole-body control](https://humanoidintel.ai/glossary/whole-body-control) for mobile manipulation is advancing rapidly. The remaining hard problem is hands — specifically, collecting enough high-quality demonstration data to train policies that generalize to the messy contact interactions of real environments.

The field's current approaches each carry the costs that DITTO identifies. Pure vision-based teleoperation systems lack force feedback and produce demonstrations that look correct but feel wrong to downstream policy learning. Glove-based in-the-wild systems produce rich force data but require sim-to-real transfer across a substantial embodiment gap. Some groups are exploring [imitation learning](https://humanoidintel.ai/glossary/imitation-learning) from video at scale, but zero-shot generalization to new contact geometries remains unreliable without grounded force information in the training distribution.

DITTO's co-design philosophy represents a different bet: instead of solving the domain gap in software (through better sim-to-real transfer or larger foundation models), close it in hardware by making the data-collection device and the deployment device kinematically identical. This is a more expensive upfront engineering commitment, but it sidesteps a class of transfer problems that have resisted purely algorithmic solutions.

The Ciocarlie lab at Columbia has a credible track record in dexterous hand research, which gives this paper more weight than a first-time entrant in the space. The authors include Matei Ciocarlie, a well-known figure in the manipulation research community, alongside a team that spans mechanical and software co-design expertise.

---

## Skeptical Analysis: What DITTO Doesn't Yet Show

The abstract is careful about its claims, and that care itself signals where the gaps are.

**Policy generalization scope is unspecified.** The paper demonstrates "learned policies on contact-rich tasks" — but the number of tasks, the diversity of objects, and the success rates are not detailed in the abstract. Whether this generalizes beyond the training distribution, and whether the learned policies transfer to a full humanoid form factor, remains open.

**Exoskeleton wearability at scale.** Motorized exoskeletons have a well-documented history of being excellent research instruments and difficult commercial products. Factors like calibration drift, inter-operator anatomical variation, and fatigue during long collection sessions are non-trivial engineering challenges that a co-design paper may bracket for future work.

**7-DOF is capable but not exceptional.** Human hands have substantially more degrees of freedom when you account for wrist, MCP, PIP, and DIP joints across all fingers. A 7-DOF system covering index-to-thumb workspace is a meaningful subset for manipulation — pinch grasps, precision grips — but it does not address whole-hand power grasps or multi-finger in-hand reorientation that some humanoid use cases require.

**Bilateral teleoperation latency.** Joint-level force feedback in bilateral teleoperation is notoriously sensitive to communication latency. The paper does not, in the abstract, specify the control loop rates or latency budget. This will matter significantly for remote operation scenarios versus co-located demonstration collection.

---

## Industry Trajectory Implications

DITTO arrives at a moment when the humanoid industry is consolidating around the recognition that data infrastructure is as important as hardware. [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) has built its entire thesis around scalable data collection for manipulation policies. [Figure AI](https://humanoidintel.ai/companies/figure-ai) has invested heavily in teleoperation infrastructure. The academic contribution here is to show that the hardware co-design layer — the interface between human demonstrator and robot learner — deserves the same systematic attention as the policy architecture layer above it.

If the full paper's experimental results support the abstract's claims, DITTO's co-design methodology is a credible template for how humanoid companies should approach next-generation hand and data-collection interface development in tandem, rather than treating them as sequential engineering problems.

---

## Key Takeaways

- **DITTO is a 7-DOF robotic hand and kinematically equivalent motorized exoskeleton**, co-designed anatomically to unify handheld data collection and bilateral teleoperation in a single platform.
- **The 1-to-1 actuator mapping** enables joint-level force feedback and eliminates the kinematic remapping layer that degrades bilateral teleoperation fidelity.
- **The core innovation is closing the embodiment gap in hardware**, not software — making the data-collection interface and the deployment hand kinematically identical.
- **The system spans the operator's natural index-to-thumb workspace** and demonstrates learned policies on contact-rich tasks.
- **Open questions** include policy generalization breadth, exoskeleton wearability at scale, and bilateral teleoperation latency performance.
- **For the humanoid industry**, this represents a hardware-first answer to the manipulation data bottleneck that every major program is currently navigating.

---

## Frequently Asked Questions

**What is DITTO in robotics?**
DITTO (Dexterous Interface for Transparent TeleOperation) is a system from Columbia University that pairs a 7-DOF robotic hand with a kinematically equivalent motorized exoskeleton. The 1-to-1 actuator mapping allows the same platform to perform handheld in-the-wild data collection and bilateral teleoperation with joint-level force feedback.

**Why is force feedback important for dexterous manipulation data collection?**
Contact-rich manipulation tasks — picking fragile objects, threading, precision assembly — depend on force modulation that operators cannot replicate without tactile feedback. Without force feedback in teleoperation, collected demonstrations lack the subtle grip adjustments that downstream learned policies need to generalize to real deployment conditions.

**What is the embodiment gap in robot learning?**
The embodiment gap refers to the visual and kinematic mismatch between a data-collection device (e.g., a human hand in a glove) and the robot that must execute the learned policy. Policies trained on human hand demonstrations must bridge differences in hand geometry, contact surface, and visual appearance — a transfer problem that DITTO addresses by making the collection device and robot hand kinematically identical.

**How does DITTO differ from standard teleoperation gloves?**
Standard instrumented gloves track human hand pose and remap it to a robot hand, which typically has a different kinematic structure. DITTO's co-design approach creates a robot hand that *is* kinematically equivalent to the exoskeleton, enabling 1-to-1 joint mapping and bilateral force reflection — capabilities that remap-based systems approximate but cannot achieve exactly.

**What humanoid companies are working on dexterous hand data collection?**
Multiple humanoid programs treat manipulation data infrastructure as a core competitive asset. [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence) focuses on scalable policy learning from diverse manipulation data. [Figure AI](https://humanoidintel.ai/companies/figure-ai) has built teleoperation infrastructure for its humanoid platform. Academic work like DITTO from the Ciocarlie lab at Columbia directly informs the hardware interface layer that all of these programs depend on.